Flux#
Eimi carries five Flux checkpoints: four Flux 2 Klein variants grouped under the Klein family entry, and one standalone Flux Dev. All of them take plain prose prompts and none of them use a negative prompt.
At a glance#
| Model | Picker group | Steps | CFG | Sampler | Scheduler | Steps range | CFG range | Base size | Batch | Negative | Model-specific LoRAs |
|---|---|---|---|---|---|---|---|---|---|---|---|
| Klein 9B Fast | Klein | 4 | 1 | euler |
simple |
2-8 | 1-2 | 1024x1024 | 2 | No | 3 |
| Klein 9B Quality | Klein | 20 | 5 | euler |
simple |
15-30 | 3-7 | 1024x1024 | 2 | No | 4 |
| Klein 4B Fast | Klein | 4 | 1 | euler |
simple |
2-8 | 1-2 | 1024x1024 | 2 | No | 0 |
| Klein 4B Quality | Klein | 20 | 5 | euler |
simple |
15-30 | 3-7 | 1024x1024 | 2 | No | 1 |
| Flux Dev | direct pick | 20 | 3.5 guidance | euler |
simple |
15-30 | 1.5-10 | 1024x1024 | 2 | No | 0 |
The picker lists the Klein versions in that order: 9B Fast, 9B Quality, 4B Fast, 4B Quality. Flux Dev sits on its own as a direct pick.
None of the five pins a default upscaler, so hi-res fix and /upscale fall back to the global default, 4x-AnimeSharp.pth.
Choosing between the four Klein variants#
The Klein entries are two axes crossed: model size (4B or 9B) and sampling recipe (Fast or Quality). Both axes cost real time, and they cost it in different ways.
| 4B Fast | 4B Quality | 9B Fast | 9B Quality | |
|---|---|---|---|---|
| Weights | Klein 4B, distilled | Klein 4B, base | Klein 9B, distilled | Klein 9B, base |
| Text encoder | Qwen3 4B | Qwen3 4B | Qwen3 8B (fp8 mixed) | Qwen3 8B (fp8 mixed) |
| Sampling work per image | 4 passes | 40 passes | 4 passes | 40 passes |
Does cfg: do anything? |
Barely -- capped at 2 | Yes, 3 to 7 | Barely -- capped at 2 | Yes, 3 to 7 |
| Model-specific LoRAs available | 0 | 1 | 3 | 4 |
| License | Apache 2.0 | Apache 2.0 | Non-commercial | Non-commercial |
Two things to take from that table:
- Fast versus Quality is roughly a tenfold difference in sampling work. Fast runs 4 steps at CFG 1, which is 4 passes through the model. Quality runs 20 steps at CFG 5, and any CFG above 1 makes the sampler evaluate the model twice per step -- so 40 passes. Quality is not "a bit slower", it is a different order of wait.
- 4B versus 9B is a difference in what has to be loaded. The 9B weights need the 8B text encoder rather than the 4B one, because the embedding dimensions differ. On a single home GPU that means more VRAM held and a longer model swap when you switch back and forth. If you are alternating between Klein and another family, the 4B entries hurt less.
A reasonable way to work
Draft on 4B Fast until the composition is right, then re-run the same prompt on 9B Quality for the final. The two share a prompt style and a resolution table, so nothing else has to change.
The models#
Klein 9B Fast#
The larger Klein weights on the 4-step distilled recipe. This is the quickest way to get a 9B-quality subject onto the card.
| Architecture | Flux 2 Klein 9B (fp8) |
| Subtype | klein_9b, klein_distilled |
| Prompting | Prose, 2 to 4 sentences |
| Negatives | None -- the graph has no negative path |
| Steps | 4 (2-8) |
| CFG | 1 (1-2) |
| Sampler | euler |
| Scheduler | simple |
| Base resolution | 1024x1024 |
| Default batch | 2 |
| Default upscaler | None pinned -- falls back to 4x-AnimeSharp.pth |
Good at: fast prose-prompt scenes with better subject coherence than 4B at the same step count, the widest LoRA support of the Fast pair (3 tagged LoRAs), quick iteration on composition.
Weak at: long or multi-clause prompts -- it is distilled, so the recommended CFG ceiling is 2 and you cannot turn guidance up to force obedience; anything you wanted excluded, because there is no negative prompt; VRAM economy, since the 9B weights drag the 8B text encoder along; commercial use, as it carries a non-commercial licence.
Klein 9B Quality#
The base 9B weights at 20 steps and CFG 5. The most capable and the slowest thing on this page.
| Architecture | Flux 2 Klein 9B base (fp8) |
| Subtype | klein_9b, klein_base |
| Prompting | Prose, 2 to 4 sentences |
| Negatives | None -- the graph has no negative path |
| Steps | 20 (15-30) |
| CFG | 5 (3-7) |
| Sampler | euler |
| Scheduler | simple |
| Base resolution | 1024x1024 |
| Default batch | 2 |
| Default upscaler | None pinned -- falls back to 4x-AnimeSharp.pth |
Good at: following long prompts with several separate requirements, fine detail, responding to cfg: in the way you would expect a normal model to; it is the only Klein entry where all four Klein-tagged LoRAs apply.
Weak at: turnaround -- 40 model passes plus the 8B encoder is the heaviest Klein job and it queues behind everything else on one GPU; negatives, which still do not exist; commercial use, same non-commercial licence as 9B Fast; batch work, since a default batch of 2 at this recipe is a long wait for two images.
Klein 4B Fast#
The compact weights on the 4-step recipe. The cheapest render in the family and the one to draft on.
| Architecture | Flux 2 Klein 4B (fp8) |
| Subtype | klein_4b, klein_distilled |
| Prompting | Prose, 2 to 4 sentences |
| Negatives | None -- the graph has no negative path |
| Steps | 4 (2-8) |
| CFG | 1 (1-2) |
| Sampler | euler |
| Scheduler | simple |
| Base resolution | 1024x1024 |
| Default batch | 2 |
| Default upscaler | None pinned -- falls back to 4x-AnimeSharp.pth |
Good at: speed, seed-hunting and composition drafts, staying out of the way of other jobs (smallest weights, smallest encoder), permissive Apache 2.0 licensing.
Weak at: detail and small features at 4 steps; prompt obedience, with the same CFG-2 ceiling as 9B Fast; LoRAs -- no LoRA in the library is tagged for it, so only the universal ones are offered; negatives, which do not exist here either.
The one Flux entry with a TIPO enhancer
Klein 4B Fast is configured to expand prompts with TIPO (a small local tag model running on the CPU inside ComfyUI) rather than the LLM path the rest of the family uses. If you pass enhance: True on this model you get a short tag-style expansion, not the prose paragraph you get from AI Enhance elsewhere. Enhancement is opt-in for every Flux model, so nothing happens unless you ask for it.
Klein 4B Quality#
The compact weights on the 20-step CFG 5 recipe. Real guidance response without the 8B encoder.
| Architecture | Flux 2 Klein 4B base (fp8) |
| Subtype | klein_4b, klein_base |
| Prompting | Prose, 2 to 4 sentences |
| Negatives | None -- the graph has no negative path |
| Steps | 20 (15-30) |
| CFG | 5 (3-7) |
| Sampler | euler |
| Scheduler | simple |
| Base resolution | 1024x1024 |
| Default batch | 2 |
| Default upscaler | None pinned -- falls back to 4x-AnimeSharp.pth |
Good at: long prompts on a modest VRAM budget, a usable cfg: lever from 3 to 7, Apache 2.0 licensing, being the cheaper of the two Quality options by a wide margin.
Weak at: raw detail and world knowledge compared to 9B Quality -- it is the same recipe on less than half the parameters; LoRAs, with exactly one tagged for it; negatives; and it is still 40 model passes, so it is not a drafting model.
Flux Dev#
The original Flux.1 Dev, repacked to fp8 with its text encoders and VAE bundled in. The photoreal option on this page, and the only one that is not a Klein.
| Architecture | Flux.1 Dev 12B rectified-flow DiT, fp8 bundled repack |
| Subtype | none |
| Prompting | Prose, 2 to 4 sentences |
| Negatives | None -- an empty string is passed to satisfy the sampler |
| Steps | 20 (15-30) |
| CFG | cfg: is the guidance value, default 3.5 (1.5-10) |
| Sampler | euler |
| Scheduler | simple |
| Base resolution | 1024x1024 |
| Default batch | 2 |
| Default upscaler | None pinned -- falls back to 4x-AnimeSharp.pth |
Good at: photorealism, materials and skin, and following an ordinary English description of a scene -- its config description calls out photorealism and strong prompt following, and that is what it is here for.
Weak at: anime and illustration styles, which the SDXL family and Anima do far better; LoRAs, since none of the 91 in the library are tagged for it, leaving only the universal ones; anything you want removed from the image, because there is no negative prompt; and it is the oldest architecture in the catalog on a non-commercial licence.
Flux Dev treats cfg: as a guidance number
Flux Dev does not use real classifier-free guidance. Its sampler CFG is pinned to 1.0 internally and the value you pass in cfg: is repurposed as the Flux guidance number instead. The useful range is 1.5 to 10, and around 3.5 is the usual starting point. Raising it does not behave like raising CFG on an SDXL model, and the job details panel will show your guidance value in the field labelled "cfg".
Prompt style#
Flux models want prose. Two to four vivid sentences, subject first.
/imagine prompt: A knight in ornate silver armor standing on a misty clifftop at dawn, cape snapping in the wind, cold blue light raking across the wet rock, shot on a long lens with a shallow depth of field model: Klein 9B Quality
Structure that works: subject, then appearance, then action, then environment, then lighting, then camera.
| Do | Don't |
|---|---|
| Describe concretely: "a chipped enamel mug on a slate counter" | Use quality tags -- masterpiece, best quality, 8k, hdr do nothing here |
| Say "with emphasis on the red scarf" | Use weight syntax -- (red scarf:1.5) is not parsed, it is just read as text |
| Name the lighting and the lens | Use vague praise words: "beautiful", "amazing", "stunning" |
| Set an actual environment | Ask for a "white background" -- it tends to blur the output on Flux Dev |
No Flux model adds quality tags for you. Your prompt is encoded exactly as written, minus any LoRA trigger words that get injected. See the prompting guide and AI Enhance, which uses a natural-language instruction set for this family.
Character names are rewritten in natural form on all five models -- Hatsune Miku, not the escaped Danbooru tag the SDXL family gets. See Character matching.
No negative prompt at all
Flux 2 Klein has no negative prompt path in its graph, and Flux Dev passes an empty one. Anything you type in negative: is discarded on all five models. Phrase what you want positively.
Resolutions#
All four Klein models and Flux Dev default to 1024x1024.
Klein:
| 1:1 | 9:7 | 7:9 | 4:3 | 3:4 | 3:2 | 2:3 | 16:9 | 9:16 | 21:9 | 9:21 |
|---|---|---|---|---|---|---|---|---|---|---|
| 1024x1024 | 1136x880 | 880x1136 | 1152x896 | 896x1152 | 1216x832 | 832x1216 | 1344x768 | 768x1344 | 1536x656 | 656x1536 |
Flux Dev:
| 1:1 | 16:9 | 9:16 | 4:3 | 3:4 | 3:2 | 2:3 | 21:9 | 9:21 |
|---|---|---|---|---|---|---|---|---|
| 1024x1024 | 1344x768 | 768x1344 | 1152x896 | 896x1152 | 1216x832 | 832x1216 | 1536x640 | 640x1536 |
Klein edit modes#
Klein is also an editing model. When you edit or restyle an image with Klein, you choose one of four modes, which combine model size with speed:
| Mode | Model | Steps | CFG |
|---|---|---|---|
| 4B Fast | Klein 4B, distilled | 4 | 1.0 |
| 4B Quality | Klein 4B, base | 20 | 5.0 |
| 9B Fast | Klein 9B, distilled | 4 | 1.0 |
| 9B Quality | Klein 9B, base | 20 | 5.0 |
The tradeoff is the same one described above: Fast is 4 model passes, Quality is 40, and the 9B modes pull in the larger text encoder.
Edit limits: steps 1 to 50, CFG 0.5 to 10.0, batch 1 or 2. You can supply one or two input images -- with two, the second acts as the style reference.
See /edit and restyle for how to launch these.
The 1 to 4 megapixel budget#
Klein edits render at the pixel count of your input image, clamped to between 1 MP and 4 MP. A small input is scaled up to 1 MP; a huge input is scaled down to 4 MP. If the input dimensions cannot be read, it defaults to 4 MP.
The resolution presets override that budget directly:
| Preset | Pixel budget |
|---|---|
| 1k | 1,048,576 px (1 MP) |
| 2k | 4,194,304 px (4 MP) |
If you also set an aspect ratio, the width and height are recomputed from the clamped budget and rounded to a multiple of 64.
Restyle strength#
Restyle strength runs 0.0 to 1.0 and controls how much of the source image survives into the result. Low values drift further from the original; high values keep its structure.
The style reference gets 75% weight
When you supply a second image as a style reference, its strength is set to 75% of the source strength automatically. You control one slider; the reference tracks it.
LoRAs and other features#
- Up to 4 LoRAs per job on all Flux models, applied to the model and both text encoders. Which ones are offered depends on the entry: 3 LoRAs are tagged for the 9B pair, 1 for the two Quality (base) entries, and none specifically for 4B Fast or Flux Dev. Universal LoRAs -- the ones with no compatibility list -- are offered on all five. See the LoRA catalog.
- Detailer works on Flux models, so faces, eyes and hands can be fixed after the fact.
- Ultimate SD Upscale runs, but it replaces your scene prompt with a generic detail prompt for every tile. That is deliberate -- Flux paints a whole miniature scene into each tile otherwise. See Upscaling methods.
- Hi-res fix works, taking its upscaler from the global default since no Flux model pins one.
Related#
- Model overview -- the full catalog
- /imagine -- every parameter
- /edit -- where the Klein edit modes live
- Upscaler catalog