/edit#
/edit is the one command that both generates and edits images. You pick a model, give a prompt, and optionally attach up to four input images. Its command description reads "Generate or edit images with your choice of AI model".
It works in servers, DMs and group DMs, and is installed both as a guild install and a user install, so it follows the usual public vs ephemeral rules.
/edit prompt: give her a red jacket image1: [attachment] model: klein quality: 9b_quality
/edit prompt: a lighthouse in a storm model: nanopro aspect_ratio: 21:9 resolution: 2K
Four ways into the same machinery#
| Entry point | What it gives you | Restyle available? | Detailer button? |
|---|---|---|---|
/edit slash command |
Everything, typed as parameters | Yes, via model: restyle |
No |
| 🎨 Edit button on a finished job card | The edit menu, pre-set to your saved edit model | No | No |
| Right-click → Apps → tweak | The same edit menu, aimed at any image in Discord | Yes | Yes |
| Right-click → Apps → restyle | Style presets only, no model picker | Always | No |
The slash command is the only path that runs the per-user queue-limit check. The menus skip it.
Parameters#
There are 13. Only prompt is required by Discord.
| Parameter | Type | Default | What it does |
|---|---|---|---|
prompt |
text (required) | -- | What you want to generate or edit, in natural language |
image1 |
attachment | none | Input image to edit |
model |
choice of 14 | klein |
Which model runs the job |
aspect_ratio |
text (autocomplete) | none | Output aspect ratio. Ignored on several models, see below |
resolution |
text (autocomplete) | none | Pixel-size tier. Only four models read it |
quality |
text (autocomplete) | per model | Render-cost knob. Options depend on the model |
google_search |
true/false | False |
Google Search grounding. nano2 and nanopro only |
thinking |
minimal or high |
minimal |
Reasoning level. nano2 and nanopro only |
batch |
number | 1 |
How many images to generate |
image2 |
attachment | none | Second input image |
image3 |
attachment | none | Third input image |
image4 |
attachment | none | Fourth input image |
style |
text (autocomplete) | none | Style preset. Required by restyle, ignored by everything else |
Autocomplete follows the model you already typed
The quality, resolution and aspect_ratio autocompletes read whichever model is currently filled in, so pick the model first and the other lists narrow to what that model accepts. The resolution list comes back empty for every local model, because those size themselves rather than offering tiers.
The style autocomplete respects your own preset ordering and hidden-style choices from /settings, and shows at most 25 entries.
prompt#
Free text. For every model except restyle it must contain at least 3 non-space characters, or you get "Please provide a more detailed edit prompt (at least 3 characters)."
There is no upper limit imposed by the bot. Discord's own 6000-character ceiling on a slash-command text option is the real cap, which is far more than any edit instruction needs. (The modal in the menu path is a different story: it caps at 2000.)
On restyle, prompt is still required by Discord but the length check is skipped, and whatever you type is appended to the generated style sentence. /edit model: restyle style: ghibli prompt: keep her hair colour produces "re-create this image in the style of Studio Ghibli films... keep her hair colour".
On z_image your prompt does double duty: it is both the mask target handed to SAM3 and the edit instruction. That is a real limitation of the slash path, and the reason the tweak menu exists for that model.
image1#
The image to edit. Must have an image content type, and must be under 25 MB. Both checks run on every attachment you pass, and the failure message names which slot broke: "Image 2 is not a valid image file." or "Image 2 is too large. Please use images under 25MB."
Six models refuse to run without it: klein, qwen, krea2, mage, restyle and z_image. Leaving it off gives "The klein model requires an input image. Attach an image to the image1 parameter."
model#
One of 14 fixed choices. Full breakdown in the next section. Picking a VIP model without VIP gives you "The Nano Banana model is available to VIP members only. Try 'qwen' or 'klein' instead." Admins skip the gate.
aspect_ratio#
Accepts 16:9, 3/4, 3 : 4, or a raw size like 1024x768, and snaps to the closest ratio the model actually supports by comparing log-ratios.
What it does depends heavily on the model, and this is the parameter most likely to do nothing:
| Model | Effect |
|---|---|
klein, restyle |
Works, but only in W:H form. 3/4 and 1024x768 are silently dropped and the edit keeps the input's shape |
qwen |
Nothing. The autocomplete offers ten ratios, the edit graph never reads the value |
krea2, mage, z_image |
Nothing. The value is not even passed along |
nano, nano2, nano2lite, nanopro |
Works. Anything outside the model's list is snapped to the closest one it allows |
gpt, gpt_15, gpt_2, gpt_1_mini |
Works, and is resolved into concrete pixel dimensions. Anything past 3:1 or 1:3 is clamped |
With no aspect_ratio at all, the ratio of image1 is detected and used. With no image and no ratio, models fall back to square.
aspect_ratio is a no-op on five of the fourteen models
Qwen, Krea 2, Mage and Z-Image ignore it outright, and Klein/Restyle only understand the W:H form. There is no error message when it is ignored -- the job runs and comes back the shape of your input. If you need a specific shape from a local model, resize the input first.
resolution#
A pixel-size tier, not a dimension. Only four models read it:
| Model | Accepts | Default |
|---|---|---|
nano2 |
0.5K, 1K, 2K, 4K |
1K |
nanopro |
1K, 2K, 4K |
1K |
gpt, gpt_2 |
1K, 2K, 4K |
1K |
nano2lite is forced to 1K whatever you type. nano, gpt_15 and gpt_1_mini ignore it (the legacy OpenAI models only have three fixed sizes). Every local model ignores it too, and its autocomplete returns nothing on those, which is the quickest way to tell.
quality#
The render-cost knob. It is not a pixel size. Options per model:
| Model | Options |
|---|---|
klein, restyle |
4b_fast, 4b_quality, 9b_fast, 9b_quality, plus the legacy bare fast / quality |
qwen, krea2, mage |
fast, quality |
gpt, gpt_15, gpt_2, gpt_1_mini |
auto, low, medium, high |
z_image |
low, medium, high -- accepted, but the autocomplete offers nothing |
nano, nano2, nano2lite, nanopro |
None. The value is ignored |
For Klein and Restyle the value carries two things at once. A value starting with 4b_ or 9b_ sets both the model size and the speed. The bare fast and quality set only the speed, and the size comes from your saved edit model in /settings. Leave it empty and both come from that setting -- and on restyle, from your saved restyle model rather than your edit model.
Leaving quality empty on qwen gives you Fast, not Quality
The capability table lists Qwen's default as quality, and the tweak menu pre-selects Fast (8 steps). The slash command falls through to the 8-step recipe as well. If you want the 40-step pass from /edit, type quality: quality explicitly.
Z-Image's denoise values are typeable but not suggested
z_image has no autocomplete entries for quality, so it looks like the parameter does nothing. Type low, medium or high by hand and they map to 20%, 40% and 70% denoise. Anything else, including an empty value, lands on 40%. The Custom denoise box only exists in the menu.
google_search#
Turns on Google Search grounding for nano2 and nanopro, so the model can pull real-world facts into what it draws. Set it on any other model and the flag is dropped before the job is built. No error.
thinking#
minimal (the default) or high, again for nano2 and nanopro only. high also turns on returning the model's thoughts. On every other model the value is discarded.
batch#
How many images come back.
| Model | Allowed |
|---|---|
klein, restyle |
1 or 2 |
qwen, krea2 |
1. Anything higher gets "Batch is not supported with the qwen model." |
mage, z_image |
1. Anything higher gets "Batch must be between 1 and 1 for Mage Flow Edit." |
nano, nano2, nano2lite, nanopro, all gpt* |
1 to 4 |
Zero and negative numbers hit the same range error. On the API models a batch of 4 costs the maintainer four separate API calls, so it is not a free knob.
image2, image3, image4#
Extra inputs for composition and reference. The ceiling is per model: klein 2, qwen 3, mage 3, everything on an API 4, and krea2 / restyle / z_image exactly 1. Going over gives "Flux2 Klein supports up to 2 image(s)." with the model's full name.
image4's help text is out of date
The parameter describes itself as "Fourth input image (optional, nano/nano2/nanopro only)". In practice nano2lite and all four GPT variants take a fourth image too. The text is stale, not the behaviour.
style#
A style preset key. restyle refuses to run without one: "The restyle model requires a style parameter. Use autocomplete to see available presets." Pass it to any other model and it is ignored.
25 presets are loaded at the moment. 23 are prompt-only; two (Anime and Omori) work by quietly attaching a reference picture as a second input image, which is why Restyle's own input limit is 1.
The 14 models#
model |
Runs on | VIP | Max inputs | Max batch | Quality options | Resolution tiers | Aspect ratios |
|---|---|---|---|---|---|---|---|
klein |
Local GPU | No | 2 | 2 | 4b_fast, 4b_quality, 9b_fast, 9b_quality |
none | Standard 10 |
qwen |
Local GPU | No | 3 | 1 | fast, quality |
none | Offered, ignored |
krea2 |
Local GPU | No | 1 | 1 | fast, quality |
none | Not offered |
mage |
Local GPU | No | 3 | 1 | fast, quality |
none | Not offered |
restyle |
Local GPU | No | 1 | 2 | 4b_fast, 4b_quality, 9b_fast, 9b_quality |
Default / 1K / 2K, menu only | Standard 10 |
z_image |
Local GPU | No | 1 | 1 | low, medium, high (+ Custom in the menu) |
none | Not offered |
nano |
Gemini API | Yes | 4 | 4 | none | none | Standard 10 |
nano2 |
Gemini API | Yes | 4 | 4 | none | 0.5K, 1K, 2K, 4K |
Standard 10 plus 1:4, 4:1, 1:8, 8:1 |
nano2lite |
Gemini API | Yes | 4 | 4 | none | 1K only, forced |
Standard 10 |
nanopro |
Gemini API | Yes | 4 | 4 | none | 1K, 2K, 4K |
Standard 10 |
gpt |
OpenAI API | Yes | 4 | 4 | auto, low, medium, high |
1K, 2K, 4K |
Standard 10 |
gpt_15 |
OpenAI API | Yes | 4 | 4 | auto, low, medium, high |
Fixed three sizes | 1:1, 3:4, 4:3, 2:3, 3:2 |
gpt_2 |
OpenAI API | Yes | 4 | 4 | auto, low, medium, high |
1K, 2K, 4K |
Standard 10 |
gpt_1_mini |
OpenAI API | Yes | 4 | 4 | auto, low, medium, high |
Fixed three sizes | 1:1, 3:4, 4:3, 2:3, 3:2 |
"Standard 10" means 1:1, 3:4, 4:3, 2:3, 3:2, 4:5, 5:4, 9:16, 16:9, 21:9.
VIP models
Every nano* and gpt* variant is VIP-only. Non-VIPs get turned away by name; admins bypass the gate. In the tweak and Edit menus, VIP models are hidden from the dropdown entirely rather than being shown and refused.
The model catalog drifts
Model entries get added, renamed and pulled without any release process. If a model: value here does not appear in autocomplete, it is gone. The six local models are the stable core; the eight API ones depend on somebody else's product decisions.
What the API models cost#
The Gemini and OpenAI variants run on the maintainer's personal API keys and are billed to them per image generated. That is the whole reason those eight are VIP-gated -- not exclusivity, just a bill. A batch: 4 on nanopro at 4K is four full-price calls in one click.
If a key runs dry, if a provider deprecates a preview model, or if the maintainer decides the spend is not worth it, those entries stop working or disappear. The six local models (klein, qwen, krea2, mage, restyle, z_image) keep running regardless, because they cost electricity rather than money.
Local models#
klein -- Flux 2 Klein#
The default. Four modes crossing model size (4B or 9B) with sampling recipe (Fast or Quality). Same weights as the Klein family in /imagine.
| Max input images | 2 |
| Max batch | 2 |
| Quality options | 4b_fast (4 steps, CFG 1.0), 4b_quality (20 steps, CFG 5.0), 9b_fast (4 steps, CFG 1.0), 9b_quality (20 steps, CFG 5.0) |
| Default quality | 4b_fast, unless your saved edit model says otherwise |
| Resolution tiers | None |
| Aspect handling | W:H ratios only. Output renders at the input image's pixel count, clamped between 1 MP and 4 MP |
| VIP | No |
Good at: general instruction edits, which is what it is the default for. Two-image work, so "put the person from image 2 into image 1" is in reach. Keeping the output the same size and shape as your input without being asked. Batch 2, so you can compare two attempts of the same instruction side by side.
Weak at: anything needing three or more references, which is Qwen or Mage territory. The 9B modes need a bigger text encoder loaded alongside them, so switching between 9B and anything else on one home GPU costs a model swap you will feel in the queue. There is no negative prompt anywhere in the graph, so exclusion has to be phrased positively. And the 20-step Quality recipes at 4 MP are a genuinely long job.
qwen -- Qwen Image Edit#
The three-reference option. Runs a 4-bit Nunchaku build of Qwen Image Edit 2509.
| Max input images | 3 |
| Max batch | 1, always |
| Quality options | fast (8 steps, CFG 1.0), quality (40 steps, CFG 4.0) |
| Default quality | fast on the slash command, despite what the tables elsewhere claim |
| Resolution tiers | None |
| Aspect handling | None. The graph never reads the value. Output follows the input image |
| VIP | No |
Good at: carrying three separate reference images into one instruction, which is more than Klein manages. Its 40-step Quality pass is the highest-effort local edit available and holds fine detail better than any 4-step recipe.
Weak at: batching, permanently. Aspect control, which it advertises and then ignores. Turnaround at 40 steps, which is five times the Fast recipe's work. And the Nunchaku node it depends on has a history of breaking on ComfyUI updates, so this is the local model most likely to be temporarily down.
krea2 -- Krea 2 Edit#
Single-image editing on Krea 2 plus an identity-preserving LoRA. Its Quality mode is specifically an object-removal recipe.
| Max input images | 1 |
| Max batch | 1 |
| Quality options | fast (Krea 2 Turbo, 10 steps, CFG 1.0), quality (Krea 2 Raw, 20 steps, CFG 3.0) |
| Default quality | fast |
| Resolution tiers | None |
| Aspect handling | Not offered at all |
| VIP | No |
Good at: removals. The Quality mode is built from the LoRA author's own "delete salient content" recipe, and at CFG 3 the base model fills the hole instead of re-drawing the thing you asked it to delete. Also good at keeping a face recognisably the same person across an edit, which is what the identity LoRA is for.
Weak at: everything involving more than one picture. No second input, no batch, no aspect control, no resolution tier. It does one thing to one image. Note also that the Fast mode is the wrong tool for removals specifically: at CFG 1 the turbo model tends to put the subject back.
mage -- Mage Flow Edit#
Microsoft's Mage-Flow-Edit, in the same family as the Mage Flow models in /imagine. MIT-licensed and small.
| Max input images | 3 |
| Max batch | 1 |
| Quality options | fast (Edit-Turbo, 4 steps, CFG 1.0), quality (RL-aligned Edit, 20 steps, CFG 5.0) |
| Default quality | fast |
| Resolution tiers | None |
| Aspect handling | Not offered. Output follows the first input, normalised to roughly 1 MP |
| VIP | No |
Good at: being quick. Four steps on a 4B model is about the cheapest real edit in the bot, and it still takes three reference images. Handy for iterating on an instruction before committing to a slow model.
Weak at: size. Everything is normalised to about 1 MP on the way in, so you cannot get a large output from Mage no matter what you feed it. No batch, no aspect control, and the 20-step Quality mode gives up the speed advantage that was the reason to pick it.
restyle -- Klein with a style preset#
Not a separate model. It builds a style-transfer prompt from a preset and hands the job to Klein.
| Max input images | 1 (a reference preset silently occupies the second slot) |
| Max batch | 2 |
| Quality options | Same four Klein modes. Empty falls back to your saved restyle model |
| Resolution tiers | Default / 1K / 2K, reachable only from the menu's 📐 toggle |
| Aspect handling | Same as Klein: W:H only |
| VIP | No |
Good at: applying a look consistently, because the style sentence is written for you rather than typed differently every time. The two reference presets get an extra instruction telling the model to copy the art style and not the reference's subject, which is the failure mode reference-image restyling usually has.
Weak at: originality within a preset, by design. On the slash path it has no Source Influence control, so it always runs at the preset's own strength (0.8 unless the preset says otherwise) -- for that dial you need the restyle menu. It also inherits every Klein limitation, including the silent aspect-ratio drop.
z_image -- SAM3 auto-mask inpaint#
The only masked-inpainting path in the bot. You name a target in words, SAM3 grounds it into a mask, and Z-Image repaints only that region with differential diffusion.
| Max input images | 1 |
| Max batch | 1 |
| Quality options | low (20% denoise), medium (40%), high (70%), plus Custom 1-100 in the menu |
| Default quality | 40% denoise |
| Resolution tiers | None. The image stays the size it arrived |
| Aspect handling | Not offered |
| VIP | No |
Good at: surgical changes. Nothing outside the mask moves, so backgrounds, framing and untouched faces come back byte-for-byte plausible instead of subtly redrawn the way a full-image edit does. Denoise is a real dial: 20% nudges, 70% replaces.
Weak at: anything SAM3 cannot find. The mask target is text grounding, so "the sword" works and "the mood" does not, and a miss produces either no change or a change in the wrong place. From the slash command your prompt is also the mask target, which is a bad fit for most instructions -- use tweak to type the two separately. No batch, no resizing, one image only.
Gemini models (VIP)#
All four take up to 4 inputs, batch up to 4, and will generate from nothing if you give them no image. Each job bills the maintainer's Google API key.
nano -- Nano Banana (Gemini 2.5 Flash Image)#
| Max input images / batch | 4 / 4 |
| Quality options | None |
| Resolution tiers | None. The API picks |
| Aspect ratios | Standard 10 |
| Extras | None |
Good at: being the simple one. No tiers, no thinking budget, no search -- pick a ratio and go. Text-to-image and 4-image composition both work.
Weak at: control. There is no way to ask it for a larger output, and no reasoning knob when a complex instruction needs one. Superseded in most respects by nano2.
nano2 -- Nano Banana 2 (Gemini 3.1 Flash Image)#
| Max input images / batch | 4 / 4 |
| Quality options | None |
| Resolution tiers | 0.5K, 1K (default), 2K, 4K |
| Aspect ratios | Standard 10 plus 1:4, 4:1, 1:8, 8:1 |
| Extras | google_search, thinking |
Good at: the widest option surface of any model here. It is the only one that will render 8:1 or 1:8, which is the practical way to get a banner or a tall strip. Search grounding lets it pull real details into an image, and thinking: high measurably helps multi-clause instructions.
Weak at: cost and predictability. 4K with thinking: high and search on, times a batch of 4, is the most expensive single click in the bot. It is also a preview model, so the endpoint can change or vanish without notice.
nano2lite -- Nano Banana 2 Lite (Gemini 3.1 Flash Lite Image)#
| Max input images / batch | 4 / 4 |
| Quality options | None |
| Resolution tiers | 1K only, forced |
| Aspect ratios | Standard 10 |
| Extras | None |
Good at: turnaround and spend. It is the cheapest and quickest of the Gemini four, which makes it the right one for iterating on wording before you spend a nanopro call on the final.
Weak at: everything nano2 has that it does not. No size tiers, no extreme ratios, no search, no thinking. Whatever you type in resolution is overwritten with 1K.
nano2lite is missing from the command's own help text
The model option's description lists "nano/nano2/nanopro" and the gpt variants, and leaves nano2lite out. It exists, it is selectable, and it works. One line of text is stale.
nanopro -- Nano Banana Pro (Gemini 3 Pro Image)#
| Max input images / batch | 4 / 4 |
| Quality options | None |
| Resolution tiers | 1K (default), 2K, 4K |
| Aspect ratios | Standard 10 |
| Extras | google_search, thinking |
Good at: the hard instructions. This is the Pro tier, and it is the model to reach for when a request has several requirements that all have to hold at once, or when the composition matters more than the turnaround.
Weak at: speed and price, in that order. It has no 0.5K tier and no extreme ratios, so nano2 still beats it on flexibility. Also a preview model.
OpenAI models (VIP)#
All four take up to 4 inputs and batch up to 4, all four expose auto / low / medium / high as quality, and all four bill the maintainer's OpenAI key per image.
gpt and gpt_2 -- GPT Image 2.0#
These two are the same model under two names. gpt is the muscle-memory alias for "the best OpenAI model available"; gpt_2 is the explicit versioned key. Both run GPT Image 2.0.
| Max input images / batch | 4 / 4 |
| Quality options | auto (default), low, medium, high |
| Resolution tiers | 1K (about 1 MP), 2K (4 MP), 4K (8 MP), capped at 3840 px on the long edge |
| Aspect ratios | Standard 10, clamped to at most 3:1 or 1:3 |
| Dimension rule | Rounded down to a multiple of 16 |
At 1:1 that works out to 1024x1024, 2000x2000 and 2816x2816 for the three tiers. Non-square ratios compute from the same pixel budget.
Good at: flexible sizing, which no other API model here offers in the same way -- you set a pixel budget and a ratio and get a real computed size rather than a fixed preset. Also the strongest of the four at following a long, literal instruction.
Weak at: extreme shapes, since anything past 3:1 gets clamped. The high quality setting at the 4K tier is expensive enough that a batch of 4 is not a casual choice.
The menu still labels gpt as 'GPT Image 1'
In the edit menu the gpt option reads GPT Image 1 with the description "OpenAI GPT Image 1 (VIP)". The model that runs is GPT Image 2.0, the same one gpt_2 uses -- and the slash command's own error messages call it "GPT Image (2.0)", contradicting the menu. The label is stale, not the behaviour. Plain GPT Image 1 is not exposed by /edit at all.
gpt_15 -- GPT Image 1.5#
| Max input images / batch | 4 / 4 |
| Quality options | auto (default), low, medium, high |
| Resolution tiers | None. Three fixed sizes: 1024x1024, 1024x1536, 1536x1024 |
| Aspect ratios | 1:1, 3:4, 4:3, 2:3, 3:2 |
Whatever ratio you ask for is matched to whichever of the three sizes is closest. resolution is ignored entirely.
Good at: being cheaper than GPT Image 2.0 for the same shape of task, when a 1024-class output is all you need.
Weak at: size and shape. Three sizes, no tiers, nothing wider than 3:2.
gpt_1_mini -- GPT Image 1 Mini#
| Max input images / batch | 4 / 4 |
| Quality options | auto (default), low, medium, high |
| Resolution tiers | None. Same three fixed sizes as gpt_15 |
| Aspect ratios | 1:1, 3:4, 4:3, 2:3, 3:2 |
Good at: being the cheapest OpenAI call available, which makes it the one to test a prompt on.
Weak at: everything the bigger variants do better. Same three fixed sizes, same narrow ratio list, and a smaller model behind them.
What the quality settings actually change#
| Model | fast |
quality |
|---|---|---|
klein / restyle |
4 steps, CFG 1.0 | 20 steps, CFG 5.0 |
qwen |
8 steps, CFG 1.0 | 40 steps, CFG 4.0 |
krea2 |
10 steps, CFG 1.0 (Turbo weights) | 20 steps, CFG 3.0 (Raw weights, the removal recipe) |
mage |
4 steps, CFG 1.0 (Turbo weights) | 20 steps, CFG 5.0 (RL-aligned weights) |
For the GPT models quality is passed straight to OpenAI and changes what they charge and how long it takes, not a step count. For Z-Image it is a denoise percentage. For the Gemini models it does nothing.
Rules checked before it queues#
These run in order, so the first one you trip is the message you get back.
| Order | Check | What it means |
|---|---|---|
| 1 | Bot access | You need access to the bot here. In a delegated server the error tells you to ask the owner to add you with /manage add |
| 2 | VIP | Required for all nano* and gpt* variants |
| 3 | Image required | klein, qwen, krea2, mage, restyle and z_image all need image1 |
| 4 | Style required | restyle refuses to run without a style |
| 5 | Unknown model | Should be impossible from autocomplete, but gives "Unknown model 'x'." |
| 6 | Input image count | Against the per-model maximum |
| 7 | Batch | qwen and krea2 reject any batch above 1 with their own message. Everything else must sit between 1 and its maximum |
| 8 | File type | Each attachment must actually be an image |
| 9 | File size | Hard 25 MB cap per input image |
| 10 | Prompt length | At least 3 non-space characters, for every model except restyle |
| 11 | Queue limit | "You have N job(s) in queue. Your limit is L." See The Queue |
Style presets are validated last, on the restyle path only: an unknown key, a preset missing its label or prompt template, or a reference preset whose picture file is missing all stop the job with their own message.
Everything /edit can do#
Instruction editing#
The main operation. Attach an image, describe the change in plain language, get the picture back changed.
/edit prompt: change the sky to a stormy sunset, keep everything else identical image1: [attachment] model: klein quality: 4b_quality
Instructions that name what to keep tend to survive better than instructions that only name what to change. Backed by Klein, Qwen, Krea 2 or Mage locally, or by any Gemini or OpenAI model if you are VIP.
SAM3 auto-mask inpainting#
Pick z_image and you get inpainting without drawing a mask. You name a target in words, SAM3 finds it and builds the mask, and Z-Image repaints only that region.
From the menu (the good path, because target and prompt are separate fields):
What to target (for masking):
her sunglassesEdit prompt (what it should become):round gold-rimmed glassesQuality: Medium Denoise (40%)
From the slash command, where the single prompt does both jobs:
/edit prompt: her sunglasses model: z_image image1: [attachment] quality: medium
Denoise picks how far the repaint drifts: 20% for subtle changes, 40% for moderate, 70% for major, or a custom 1-100 value from the menu.
Object removal#
Handled as an instruction edit, with Krea 2's Quality mode as the tool built for it.
/edit prompt: remove the street sign completely, fill the space with the wall behind it model: krea2 quality: quality image1: [attachment]
The menu describes that option as "Raw, 20 steps CFG 3 (removals)". The Fast mode at CFG 1 tends to re-draw whatever you asked it to remove, so pick Quality for this specifically. For a removal that must not disturb the rest of the frame, z_image at high denoise with the object named as the mask target is the alternative.
Style transfer#
/edit prompt: keep her expression model: restyle style: ghibli image1: [attachment] quality: 9b_quality batch: 2
Eimi writes the style sentence for you from the preset and appends whatever you typed in prompt. 25 presets are set up at the moment. Two of them, Anime and Omori, work by attaching a reference picture as a second input and telling the model to copy only the art style. Full list on the Style Presets page.
For the Source Influence dial and a live preview thumbnail, use the restyle right-click menu instead.
Multi-image composition#
Attach two or more images and describe how they combine.
/edit prompt: put the character from image 1 into the room from image 2, matching the lighting model: qwen image1: [attachment] image2: [attachment] quality: quality
Referring to the pictures by number in the prompt is worth doing -- the models read the images in the order you attached them. Ceilings: klein 2, qwen 3, mage 3, every API model 4, and krea2 / restyle / z_image 1.
Text-to-image#
Run /edit with no attachment at all and a nano* or gpt* model.
/edit prompt: an isometric cutaway of a submarine galley, technical illustration style model: gpt_2 aspect_ratio: 4:3 resolution: 2K quality: high
Those eight are the only models that generate from nothing. For local text-to-image use /imagine or /turbo instead, which have the full model catalog, LoRAs and parameter controls that /edit does not.
Detail fixes#
The Detailer panel repairs faces, eyes, hands and whole people. Reach it from the 🧬 Detailer button on a finished job card, or from the Detailer button inside the tweak menu. Each fix is a toggle: 🧍 Person, 😊 Face, 👁️ Eyes, 🖐️ Hands. Turn on what you need, click ▶️ Run. Full detail on the Detailer page.
Hi-res fix#
Also in the Detailer panel: a toggle that upscales and re-diffuses. Two styles exist, refine (single-pass, the default) and usdu (tiled diffusion). Defaults are 1.5x scale, denoise 0.35, upscaler 4x-AnimeSharp.pth. With it on you can override the model, sampler, scheduler, upscaler, steps, CFG and prompt, each of which otherwise inherits from the parent job. See Hi-Res Fix and Detailer and Upscalers.
Background removal#
A dropdown at the bottom of the Detailer panel, placeholder "Background Removal (optional)":
| Option | Description shown |
|---|---|
| None | -- |
| 🎨 ToonOut (Anime) | Best for anime |
| 🌟 InSPyReNet | Fast, accurate |
| ⚖️ BiRefNet General | Balanced |
| 🔬 BiRefNet HR | High resolution |
| ✨ RMBG 2.0 | High quality |
Clicking ▶️ Run with nothing selected anywhere in the panel gives "Please select at least one option."
Outpainting is not supported
Eimi has no outpaint operation. Nothing in /edit, the Detailer, or any right-click menu extends a picture beyond its original borders. Changing aspect_ratio on an edit is the closest available behaviour, and that re-renders the whole image at a different shape rather than adding new canvas at the edges. If you need a wider frame, generate wide from the start with /imagine.
The edit menu, step by step#
The same menu opens from the 🎨 Edit button on a finished job card and from the tweak right-click command. The header reads Edit Image from the button and Tweak Image from the context menu; everything below it is identical apart from two differences noted as they come up.
It times out after 300 seconds, at which point the message is replaced with "⏱️ Selection timed out." Only the person who opened it can touch it -- anyone else gets "This is not your edit menu."
1. Pick a model#
A dropdown with the placeholder "Choose editing model...". It opens pre-set to whichever model your saved edit model in /settings maps to. VIP models are hidden entirely if you are not VIP, and Restyle only appears on the tweak path (the job-card Edit button omits it).
| Option | Emoji | Description shown |
|---|---|---|
| Klein | ⚡ or ✨ | Flux2 Klein 4B Fast / 4B Quality / 9B Fast / 9B Quality, mirroring your saved setting |
| Qwen Edit | 🎨 | Qwen Edit, 8 steps |
| Krea 2 Edit | 🌊 | Krea 2 Edit |
| Mage Flow Edit | 🪄 | Mage Flow Edit |
| Nano Banana | 🍌 | Gemini 2.5 Flash (VIP) |
| Nano Banana 2 | 🍌 | Gemini 3.1 Flash (VIP) |
| Nano Banana 2 Lite | 🍌 | Gemini 3.1 Flash Lite - fast (VIP) |
| Nano Banana Pro | 🍌 | Gemini 3 Pro (VIP) |
| GPT Image 1 | 🤖 | OpenAI GPT Image 1 (VIP) |
| GPT Image 1.5 | 🤖 | OpenAI GPT Image 1.5 (VIP) |
| GPT Image 2.0 | 🤖 | OpenAI GPT Image 2.0 (VIP) |
| GPT Image 1 Mini | 🤖 | OpenAI GPT Image 1 Mini (VIP) |
| Z-Image Inpaint | 🖌️ | SAM3 auto-mask + Z-Image inpaint |
| Restyle | 🎭 | Restyle with style presets |
Klein's emoji tracks your saved mode: ⚡ for the two Fast modes, ✨ for the two Quality ones.
2. Pick an image#
Only appears when the source has more than one image. A second dropdown, "Choose which image to edit...", listing up to 25 entries. A batch from a job card contributes at most 9.
3. Set the model's options#
Everything below the dropdown is rebuilt for whichever model you picked. Switching model resets quality, aspect ratio, batch, source influence, resolution, Google Search, thinking and style back to that model's defaults, so change the model first and the options after.
| Model | What appears |
|---|---|
| Klein, Restyle | One button row: Fast / Quality / 4B / 9B / Batch: 1. The active speed and size turn blue; Batch turns green at 2 and toggles between 1 and 2 |
| Qwen | Two buttons: Fast (8 steps) / Quality (40 steps), blue when active. Pre-set to Fast |
| Krea 2 | "Quality / Speed..." dropdown: Fast (10 steps) "Turbo, 10 steps CFG 1" / Quality (20 steps) "Raw, 20 steps CFG 3 (removals)" |
| Mage | "Quality / Speed..." dropdown: Fast (4 steps) "Turbo, 4 steps CFG 1" / Quality (20 steps) "RL-aligned, 20 steps CFG 5" |
| All four GPT | "Quality / Speed..." dropdown: Auto "Balanced quality and speed" / Low "Fastest, lowest cost" / Medium "Balanced" / High "Best quality, slowest" |
| Z-Image | "Quality / Speed..." dropdown: Low Denoise "Subtle changes (20%)" / Medium Denoise "Moderate edits (40%)" / High Denoise "Major changes (70%)" / Custom "Set your own denoise (1-100)" |
| nano2, nanopro, GPT Image 2.0 | "Resolution..." dropdown, pre-selected to that model's default tier so you do not have to click it |
| nano2, nanopro | 🔍 Google Search: Off and 🧠 Thinking: Minimal, side by side. Each turns green when switched on |
| Anything with aspect support | "Aspect Ratio..." dropdown, first entry Auto ("Match input image"). Absent for Krea 2, Mage and Z-Image |
| nano, nano2, nano2lite, nanopro, all GPT | "Batch size..." dropdown, "1 image" through "4 images" |
Restyle adds a block of its own: a "Choose style preset..." dropdown whose first entry is always ✍️ Custom ("Describe your own style"), then 24 presets per page. With 25 presets loaded that means two pages, so a ◀ / Styles 1/2 / ▶ row appears under it. Below that a "Source Influence..." dropdown running 30% to 100% (80% is marked Default; anything lower says "Less structure", higher says "More structure"), and a 📐 Resolution: Default button that cycles Default → 1K → 2K and turns green once it leaves Default.
4. Read the status line#
A small grey line above the buttons summarising the current selection, joined by ·. It shows, in order and only when set: Model, Quality, Style, AR (hidden on Auto), Batch (hidden at 1), Source, Res, Search.
Model: Restyle · Quality: 9b_quality · Style: Studio Ghibli · Batch: 2 · Source: 70% · Res: 1K
5. Submit#
Up to four buttons sit at the bottom:
| Button | Colour | When it appears | What it does |
|---|---|---|---|
| ✏️ Edit Prompt | Blue | Always | Opens the prompt modal |
| ▶ Submit | Green | Only when a real style preset is picked | Skips the modal and runs the generated style prompt as-is |
| 🧬 Detailer | Green | Only on the tweak path | Opens the Detailer panel for the selected image |
| Cancel | Grey | Always | Deletes the menu message |
Both submit buttons are guarded against double-clicks: the second click is swallowed rather than queuing a second job.
The prompt modal#
Titled "Edit with {model}" using the model's menu label. One paragraph field, Prompt, capped at 2000 characters.
| Situation | Pre-filled with | Placeholder | Required? |
|---|---|---|---|
| Normal edit | Nothing | "Describe what you want to do with the image..." | Yes |
| A real style preset | The generated restyle prompt | "Edit the style prompt or leave as-is..." | No |
| ✍️ Custom style | "re-create this image in the style of " | "Describe the style you want to apply..." | Yes |
Leaving a required field empty gives "❌ Prompt cannot be empty."
The Z-Image Inpaint Edit modal#
Picking Z-Image Inpaint replaces the normal modal with a three-field one titled "Z-Image Inpaint Edit".
| Field | Required | Limit | Placeholder |
|---|---|---|---|
| What to target (for masking) | Yes | 200 characters | "e.g. face, hands, background, the sword..." |
| Edit prompt (what it should become) | Yes | 2000 characters | "e.g. a smiling face with blue eyes..." |
| Custom denoise (1-100) | Only when quality is Custom | 3 characters | "e.g. 50" |
Empty fields give "Mask target cannot be empty." or "Edit prompt cannot be empty." A non-numeric denoise gives "Denoise must be a number between 1 and 100." A number outside 1-100 is clamped into range rather than rejected.
The Detailer panel#
Reached from the 🧬 Detailer button on the tweak path, or from the button on a job card. Header Detailer, purple.
| Section | Contents |
|---|---|
| Image select | Only on a batch. "Choose which image to detail...", up to 9 entries |
| Detail Fixes | 🧍 Person, 😊 Face, 👁️ Eyes, 🖐️ Hands. Green when on |
| Hi-Res Fix | A status line reading 🟢 or ⚫ with refine \| 1.5x \| denoise 0.35, and a Toggle button beside it |
| Hi-res advanced | Only when the toggle is on: 🎨 Model: Inherit, Sampler, Scheduler, Upscaler (all defaulting to "Inherit from parent job"), a ⚙ settings button showing denoise 0.35 · 1.5x, and a ✏ Prompt button |
| Background Removal | The six-option dropdown listed above |
| Status | Selected: face, eyes + hires 1.5x + bg removal |
| Actions | ▶️ Run (green), ✖️ Cancel (grey) |
Both fix sections vanish on a 4x image
Once an image has been upgraded 4x, the Detail Fixes row and the whole Hi-Res Fix section are hidden. Background removal stays available.
If you close a modal without submitting
The menu will not re-open the modal on its own, and its double-click guard has already fired. Re-run /edit or the right-click command to start again.
The restyle menu#
The restyle right-click command opens a different, simpler view titled Re-style Image with no model picker -- it is always Klein. It adds a live preview thumbnail of the selected preset, and it puts ▶ Submit before ✏️ Edit Prompt, because running the preset unchanged is the common case. Its modal is titled "Re-style Image". If you have no default style set, it pre-selects the first preset in your list so Submit works immediately. Anyone else clicking it gets "This is not your restyle menu."
Troubleshooting#
Things /edit says, and what each one means.
| Message | What happened |
|---|---|
"You don't have access to use the bot in this server. Ask the server owner to add you with /manage add" |
Delegated server, you are not on its list. See Access |
| "You don't have access to use the bot in DMs." | Same, in a DM |
| "Access denied." | Same, in an ordinary server |
| "The Nano Banana model is available to VIP members only. Try 'qwen' or 'klein' instead." | A VIP model without VIP. The name shown is the model's full name, so gpt reports itself as "GPT Image (2.0)" |
"The klein model requires an input image. Attach an image to the image1 parameter." |
One of the six local models with nothing attached |
"The restyle model requires a style parameter. Use autocomplete to see available presets." |
restyle with no preset |
| "Flux2 Klein supports up to 2 image(s)." | Too many attachments for that model |
| "Batch is not supported with the qwen model." | qwen or krea2 with batch above 1 |
| "Batch must be between 1 and 2 for Flux2 Klein." | batch outside the model's range |
| "Image 2 is not a valid image file." | That attachment is not an image |
| "Image 2 is too large. Please use images under 25MB." | Over the 25 MB per-file cap |
| "Please provide a more detailed edit prompt (at least 3 characters)." | Prompt under 3 non-space characters |
| "You have 3 job(s) in queue. Your limit is 3." | Wait for one to finish. See The Queue |
| "Unknown style preset: x. Use autocomplete to see available presets." | The style key does not exist. Presets get renamed |
| "⚠️ Style preset 'x' is misconfigured: ..." | The preset exists but is broken. Not something you can fix -- tell the maintainer |
| "⚠️ Style reference image not found for 'Anime'." | A reference preset whose picture file is missing on the machine |
| "ComfyUI client not available." | The image backend is not up. Nothing to do but wait |
| "Workflow system not available." | Same category of problem |
| "Failed to upload image 1." | The image reached Discord but not the backend. Retry |
| "⚠️ Failed to prepare Klein edit workflow." | The job could not be built. Usually a backend or model-file problem |
| "Failed to process Klein edit. Please try again." | Something broke while queueing. Retry once, then assume it is down |
| "This is not your edit menu." | Someone else's menu |
| "⏱️ Selection timed out." | The menu sat untouched for 300 seconds |
| "❌ Prompt cannot be empty." | Required modal field left blank |
| "Denoise must be a number between 1 and 100." | Non-numeric custom denoise |
| "Please select at least one option." | ▶️ Run in the Detailer with nothing toggled |
See When Things Fail for what happens after a job has already been queued.
Which model to reach for#
| What you want | Pick | Why |
|---|---|---|
| A general edit, no strong opinion | klein |
The default, two sizes, two speeds, keeps your input's shape |
| The best local edit quality | qwen with quality: quality |
40 steps is the heaviest local recipe |
| Three reference images | qwen or mage |
The only local models that take three |
| Remove an object | krea2 with quality: quality |
The 20-step Raw recipe is built for removals |
| Change one region and nothing else | z_image |
Masked inpainting, everything outside the mask is untouched |
| The quickest local edit | mage with quality: fast |
4 steps on a 4B model |
| Apply a consistent look | restyle with a style |
The style sentence is written for you |
| Four reference images | any nano* or gpt* |
Nothing local takes four |
| Generate from nothing | any nano* or gpt*, or use /imagine |
No local /edit model does text-to-image |
| An 8:1 banner or a 1:8 strip | nano2 |
The only model that offers those ratios |
| A hard multi-clause instruction | nanopro with thinking: high |
Highest-reasoning option available |
| Real-world facts in the picture | nano2 or nanopro with google_search: True |
The only two with search grounding |
| A specific pixel budget and ratio | gpt_2 |
Computes real dimensions from a tier plus a ratio |
| To test wording before spending | nano2lite or gpt_1_mini |
The two cheapest API calls |
| Fix a face, eyes or hands | The 🧬 Detailer, not a model | Targeted repair passes |
| Make an image bigger | /upscale or the hi-res fix | /edit is not an upscaler |
| Extend the canvas | Nothing. It is not supported | Generate at the shape you want from the start |
See also#
- tweak -- the same edit menu, aimed at any message
- restyle -- style presets only, with previews and Source Influence
- Style Presets -- all 25 presets
- Detailer -- face, eye, hand, hi-res and background-removal passes
- Flux -- the Klein weights behind
kleinandrestyle - Krea 2 and Z-Image -- the models behind
krea2andz_image - /imagine -- local text-to-image with the full catalog