Remix#
🎲 Remix opens a private editor for an existing job. You get every knob the original generation used, pre-filled, and you change what you want before pressing Generate. It is the fastest way to say "same thing, but 40 steps and a different sampler".
The menu is ephemeral (only you see it) and lives for 15 minutes.
Opening it#
Click 🎲 on any result card. It is available on every job type except video.
Note
If you click 🎲 on an upscale, upgrade, detailer or high-effort result, the menu opens against the original parent job's settings, not the post-processing pass. You almost always want to remix the concept, not the upscaler. The new generation still replies to the message you actually clicked.
The main dashboard#
A teal panel headed Remix {job type} #{job id}, with your prompt, the negative prompt if
there is one, and the current parameter line beneath it. The job type shown is one of:
imagine, img2img, turbo, nano, gpt, edit, klein edit, nano edit,
gpt edit, compose, funkify, hidden object.
| Section | Control | Shown when |
|---|---|---|
| Model | Change 🎨 -- opens the Model submenu | Non-edit jobs run through ComfyUI |
| Edit model | An ✏️ Edit Model dropdown | The original job was an edit |
| LoRAs | Configure ✨ -- opens the LoRA submenu | Non-edit jobs run through ComfyUI |
| Prompt | Edit 📝 | Always |
| Negative | Edit 🚫 | Always |
| Core row | Batch: n 📦 | ComfyUI non-edit jobs |
| Core row | Aspect: W:H 📐 | Always |
| Core row | Seed: n or Seed: 🎲 | ComfyUI non-edit jobs |
| Input images | A gallery of up to 10 current input images | Always (gallery only if images exist) |
| Input images | Replace / Upload 🖼️ | Always |
| Input images | Clear 🗑️ (red) | Only when images are attached |
| Denoise | Denoise: v | Only when an input image is set, ComfyUI non-edit |
| Submenus | Quality ⚙️, Sampling 🎨, Flags 🚩 | ComfyUI non-edit jobs |
| Footer | Reset ↩️ | Always |
| Footer | Generate ✅ (green) | Always |
Music (compose) jobs get a stripped dashboard -- the prompt section and the footer, nothing else. Nano and GPT jobs hide Model, LoRAs, Quality, Sampling and Flags, because those are API models with no ComfyUI graph behind them.
The colour convention#
Green means you changed it. Every control on every screen of the remix tree turns green the moment its value differs from the original job's, and goes back to grey if you set it back. The parameter line in the header updates live too, so a steps value you just overrode starts showing up there immediately.
Reset and Generate#
Reset ↩️ restores every value to the original job's. It is grey and disabled while nothing has been changed, and turns red once anything differs.
Generate ✅ queues the job. See What Generate actually does.
Edit model dropdown#
Edit jobs get a dropdown instead of the model submenu:
| Option | Model | Access |
|---|---|---|
| ⚡ Klein | Flux2 Klein, fast inpainting | Everyone |
| 🎨 Qwen Edit | Qwen Edit model | Everyone |
| 🌊 Krea 2 | Krea 2 identity-edit LoRA | Everyone |
| 🍌 Nano | Gemini 2.5 Flash | VIP |
| 🍌 Nano 2 | Gemini 3 Pro | VIP |
| 🍌 Nano 2 Lite | Gemini 3.1 Flash Lite, fast | VIP |
Switching model here clears the old model's job flags, so you can turn a Klein edit into a Nano edit in one click. Moving away from Klein also drops any restyle preset that was attached.
Submenus#
Model#
Headed Remix: Model with the current model's nickname underneath. Uses the standard
family-grouped picker: a 🎨 Model family dropdown, version buttons for whichever family
is expanded, and page ◀ / ▶ arrows. The family of the model you are currently using is
expanded automatically when you open the screen. Back ◀ returns to the dashboard.
See Models for what each family does.
LoRAs#
Headed Remix: LoRAs -- up to 4 LoRAs, weights -2.0 to 2.0. Active slots are listed as
1. Nickname (Variant) @ 0.8, or "no LoRAs active".
| Control | Behaviour |
|---|---|
| ✨ Pick LoRAs (up to 4) | Multi-select. Only LoRAs compatible with the currently selected model are offered. Each LoRA gets a base row plus one row per trigger-carrying variant, so a variant is picked like any other entry. LoRAs you already have active with variants float to the top. |
| ◀ / page / ▶ | Pagination. Pages are packed so a LoRA's variants never split across two pages. The middle counter is a label, not a button. |
| Edit Weights ⚙️ (blurple) | Opens the weights modal. Only present when at least one slot is filled. |
| Clear All 🗑️ (red) | Empties all four slots. Only present when slots are filled. |
| Back ◀ | Returns to the dashboard. |
If nothing is compatible with the selected model, the screen says "No compatible LoRAs". See LoRAs.
Quality#
Two rows: Steps: n 🔢 (1 to 100) and CFG: v 🎚️ (1.0 to 20.0), each opening a modal.
Sampling#
Two dropdowns, Sampler and Scheduler, both populated live from ComfyUI. If the live list can't be fetched they fall back to:
- Samplers: euler, euler_ancestral, heun, dpm_2, dpm_2_ancestral, lms, dpmpp_2s_ancestral, dpmpp_sde, dpmpp_2m, dpmpp_2m_sde, dpmpp_3m_sde, ddim, uni_pc, uni_pc_bh2
- Schedulers: normal, karras, exponential, sgm_uniform, simple, ddim_uniform, beta
Flags#
Three toggles, each with an ON / OFF button that is green when on:
| Flag | What it does |
|---|---|
| NSFW | Allow explicit content |
| RAW | Skip prompt enrichment and use your text verbatim |
| Enhance | Let AI expand the prompt. Turning it off also discards the stored enhanced text, so the next run uses your words. |
The modals#
| Modal | Field | Accepts | Notes |
|---|---|---|---|
| Edit Prompt | Prompt (paragraph) |
Required, 2000 chars | Pre-filled with the prompt currently in use |
| Edit Negative Prompt | Negative Prompt (paragraph) |
Optional, 2000 chars | Leave empty to clear the negative entirely |
| Edit Steps | Steps |
Integer, clamped 1 to 100 | Placeholder shows the default, 25. Non-numeric gives "❌ Invalid steps value." |
| Edit Cfg | Cfg |
Float, clamped 1.0 to 20.0 | Placeholder shows the default, 5.0 |
| Edit Denoise | Denoise |
Float, clamped 0.0 to 1.0 | Reached from the Quality path, default 1.0 |
| Set Denoise Strength | Denoise (0.0-1.0, lower = keep more original) |
Required, clamped 0.0 to 1.0 | The dashboard's Denoise button for img2img |
| Edit Seed | Seed |
Optional, 20 chars | Empty or -1 means random. A number pins the seed so Generate keeps it. Non-numeric gives "❌ Seed must be a number." |
| Edit Batch Size | Batch Size |
1 to 4 | Empty keeps the original batch size |
| Edit Aspect Ratio | Aspect Ratio |
W:H, both sides above 0 |
For example 3:4 or 16:9. Empty keeps the original |
| Edit LoRA Weights | One field per active slot, labelled 1. Nickname |
-2.0 to 2.0 each | Up to 5 fields. A blank field leaves that LoRA unchanged. Errors name the slot: "LoRA 2: must be between -2.0 and 2.0" |
| Upload Images (up to 4) | Upload Image(s) file picker |
Up to 4 files for nano and GPT jobs, 1 otherwise | Uploading nothing restores the original job's images |
| Set Input Image (img2img) | Image URL |
500 chars, must start with http:// or https:// |
The fallback when file upload is unavailable. Empty keeps the original |
What Generate actually does#
Pressing ✅ Generate does more than replay your edits.
| Behaviour | Detail |
|---|---|
| Fresh generation, always | Derivative flags from the original are stripped: upscale, detailer, upgrade, high-effort, upgrade level and parent links all go. A remix of an 8x upscale is a brand new image at normal size, not another upscale. |
| New seed | A fresh random seed is drawn unless you pinned one in the Seed modal. |
| Your settings, not theirs | Your current /settings are re-read for webp quality, inline prompts, character detection and similarity threshold. Remixing someone else's job uses your preferences. |
| Your job limit | The job is queued under your name and counts against your queue limit. |
| Re-enhancement | With the Enhance flag on, the stored enhanced prompt is cleared so it is enhanced again from scratch -- expect a different expansion than the original got. |
| Model swap cleanup | If you changed the model, any steps, CFG or denoise values that were still at the old model's defaults are dropped so the new model's defaults apply. |
| Model-specific stripping | Flux2 and Krea 2 drop carried-over negatives and LoRAs. Krea 2 additionally drops all input images, since it is text-to-image only. |
| Where it lands | The queue card replies to the message you clicked 🎲 on. The remix menu deletes itself. |
Access and your queue limit are re-checked at this point, so a menu you left open for ten minutes cannot be used to slip past a limit you have since hit.
Conjure: remix from a description#
describe results carry their own 🎲 button, one per generated description. Clicking it
opens the same interface, headed Remix (from /describe · {style}). The bot calls it
Remix on purpose -- there is only one menu to learn.
Everything on the dashboard works identically, with these differences:
| Difference | Detail |
|---|---|
| Starting point | The chosen description becomes the prompt, and the model is the one suggested for that description's style. Every control starts grey, because nothing has been changed yet. |
| Your settings applied | Default negative, default positive, NSFW, RAW, upscaler and embeddings come from your /settings, the same way /imagine picks them up. LoRAs are deliberately not auto-loaded, since the suggested model may not match your defaults. |
| Aspect ratio | The closest ratio the suggested model supports to the source image's shape, then your /settings aspect ratio, then the model's own default. |
| Extra control | A Source image: ON (img2img) / Source image: OFF (text only) 🖼️ toggle. |
| Back buttons | Every submenu's Back returns to the Conjure dashboard, not a bare remix view. |
The source image toggle decides whether the picture you described is fed back in as an img2img input, or whether the description is generated from scratch as text-to-image. Turning it on attaches the described image; turning it off removes it completely.
Note
The whole input-image section, toggle included, disappears when the selected model is text-to-image only: Flux2, Krea 2, Flux Dev, ERNIE Image and Ideogram 4. Switch to another model in the picker and it comes straight back.
Full details of the describe flow are on describe.