FLUX 3 Image Klein. A 4K Still Is Not a Clip
Black Forest Labs released FLUX 3 Image Klein on 15 October 2026: 5.6B open weights for stills, edits, and up to 10 refs. A 4K still is not a clip. Lock frame one, then image-to-video.
FLUX 3 Image Klein. A 4K Still Is Not a Clip
Black Forest Labs said on 15 October 2026 that FLUX 3 Image Klein is out as open weights: 5.6 billion parameters, text-to-image, image editing, and multi-reference generation with up to 10 reference images. The same post says it shares core still-image jobs with the full FLUX 3 Image model — native 4K, pixel-perfect editing, bounding boxes that place elements or confine an edit — and that a base checkpoint plus a step-distilled checkpoint shipped the same day, with FP8, INT8, and NVFP4 builds. ClipCanva is not affiliated with Black Forest Labs, Ideogram, VEED, Kapwing, Canva, or NVIDIA. Klein is a still desk. Ten references are a layout stack, not a signed pack shot and not an 8-second Short. Lock the line on the AI script generator. Steal shot language from prompt ideas. Lock the still, then animate it on image to video or generate the scene on the AI video generator. Do not brief “open weights, 4K, ten refs, so the commercial writes itself.”
Launch facts (confirmed 19 October 2026)
| Source | Public claim | Workflow implication |
|---|---|---|
| BFL blog: FLUX 3 Image Klein (15 Oct 2026, 4 min read) | “Today we are releasing FLUX 3 Image Klein, a 5.6B parameter open weights model for text-to-image generation, image editing and multi-reference generation with up to 10 reference images.” Shares native 4K, pixel-perfect editing, and bounding boxes with full FLUX 3 Image. Base model for fine-tuning plus a step-distilled model for low-latency generation. FP8, INT8, and NVFP4 builds “developed with NVIDIA support,” under the FLUX Kommunity License. | Launch day is 15 October. This is an image model. Ten refs are stills, not keyframes. |
| Same post | Internal evaluations: Klein “ranks first in text-to-image, multi-reference generation and image editing, ahead of larger open weights models,” and “rates above the closed models in the comparison, Ideogram V4.5 and FLUX.2 [pro].” Versus FLUX.2 [klein] 9B: “91 Elo in text-to-image and 70 in image editing, with 38% fewer parameters.” Parameter counts exclude text encoders. Krea 2 “does not support image editing.” | Those ranks are BFL’s internal evals, not a third-party bake-off you can file. Do not put “beats Ideogram” in the VO. |
| Same post | Generator weights: 11.2 GB in BF16, 5.6 GB in FP8. INT8 “runs on NVIDIA GeForce RTX GPUs with as little as 8GB of VRAM, through ComfyUI or Diffusers.” Distilled INT8: about 4 seconds for a 1MP image on an RTX 5090, under 9 seconds on an RTX 4070, measured by NVIDIA in ComfyUI. On NVIDIA GB300, distilled NVFP4: 1MP “in under 0.6 seconds.” Serving a LoRA “adds around 5% to generation time.” | Local speed is a still number. It is not time-to-an-approved-clip. |
| Same post | BFL API: distilled model as flux-3-image-klein, with prompt upsampling and pixel restoration that “preserves unedited regions of an image exactly.” “Pricing starts at $0.009 per 1K image.” Weights on Hugging Face (base and distilled). Day-one support named: ComfyUI, Hugging Face Diffusers, TensorRT-LLM, SGLang Diffusion, vLLM-Omni. LoRA: Ostris AI Toolkit and NVIDIA NeMo AutoModel recipes. |
API ID is an image endpoint. Pixel restoration is an edit guarantee, not motion. |
| Same post | FLUX Kommunity License: “free commercial use for individuals and companies with annual revenue under $5M,” plus research and non-commercial use. Larger orgs: self-hosted licensing or API. Second open-weights FLUX 3 family release after FLUX 3 Action. Safety: internal red team plus external partner Cinder. | Under $5M is a license fact, not a creative brief. FLUX 3 Action is robot control, not your pack shot. |
| BFL model page: FLUX 3 Image Klein | “Frontier quality with the speed and customization of open weights.” “One model, every task: generation, editing and multi-reference in a single set of weights.” Bounding-box examples place type, a moon, a cat. Use cases listed: photography, illustration, type and design. | “Every task” on this page is still image tasks. Type-in-the-box is not an end card with audio. |
| BFL model page: FLUX 3 Image | Separate URL. Hero: “Maximum control over every pixel.” | Klein ≠ the full FLUX 3 Image row. Do not mix IDs. |
| VEED AI Image to Video | Own URL. Nav still lists Image to Video next to AI Script Generator and AI Video Generator. | Writer, still, and renderer stay different URLs. |
| Kapwing Image To Video | Own URL. Nav still lists Script Generator next to Video Generator and Image To Video. | Same split. Do not merge them because Klein can take ten stills. |
BFL’s 15 October post is the launch-date source, the API ID source, and the license source. BFL’s own nav still splits FLUX 3 Image Klein from FLUX 3 Video. Keep that split on your desk.
What 15 October shipped — and what it did not
Operators will collapse four objects into “FLUX 3 dropped, dump the brand zip, ship the 4K hero as a Reel.” The public pages refuse that.
| Surface | What the public page says it does | What it does not do | Fail condition |
|---|---|---|---|
| 15 October Klein release | Put a 5.6B open-weights image model on the calendar | Sign the VO; pick frame one of a clip; grant a video endpoint | Treating “open weights today” as “the cut is approved” |
| Up to 10 reference images | Mix stills into one generation or edit | Invent the legal line; keep a SKU unwarped unless that still is frame one | Uploading a moodboard and calling it the pack shot |
| Bounding boxes | Place type or confine an edit to a region | Store Legal’s signed sentence; time an 8-second card | Boxing “SALE” onto a still and calling it the end card |
| Native 4K / pixel-perfect edit / pixel restoration | High-res stills; unedited pixels stay put on the API path | Motion, lip sync, native audio, a 5-second take | Exporting a 4K PNG and filing it as the Short |
| Distilled vs base | Fast 1MP stills vs a fine-tune checkpoint | Decide which take is on-model for video | Fine-tuning overnight, then generating B-roll with no locked still |
| Kommunity License / API at $0.009 per 1K | A cost and rights path for images | A ClipCanva selector row; a video credit price | Promising a client “we have FLUX 3 on the site” |
| FLUX 3 Action (prior open-weights sibling) | 7B robot-control weights, per the Klein post | Product photography | Pasting Action into an image prompt because it says FLUX 3 |
BFL’s useful split is the one to keep: Klein is generation, editing, and multi-reference in one set of image weights. That is not FLUX 3 Video. It is not a script. It is not ClipCanva’s live image-to-video selector, which still starts from one reference image.
Pick the desk before you fill 10 slots
The fail is treating ten references as “the brief is in the zip.” Map the job.
| Job | Where it actually runs | What you must supply | What the system may invent | Fail condition |
|---|---|---|---|---|
| Line a viewer will hear or read | AI script generator | Audience, seconds, the signed sentence | A second CTA, a ranking claim (“beats Ideogram”), a 4K promise | Letting Klein’s type boxes write the VO because the stamp demo has kanji |
| Shot language | Prompt ideas | Four beats, not a 10-image collage | A fifth scene, a second SKU | Dumping BFL’s photography/illustration/type list into one video prompt |
| Still that must stay on-model | Klein (or your current still model), then image to video | Pack shot as frame one | Warped type, a second flavor, a face that is not the talent | Using nine “style refs” as if they preserve the nutrition panel |
| Motion from a locked still | Image to video or the AI video generator | Locked still and/or locked caption, crop, seconds | Extra chrome, mouthed words you did not approve | Generating B-roll first, locking the still never |
| Current FLUX row on ClipCanva | FLUX 3 model page | Whatever the live page actually says | A Klein badge | Assuming ClipCanva shipped flux-3-image-klein because Hugging Face did |
| Compare after the still exists | Compare | The job (still vs clip vs script) | A winner named after a blog Elo chart | Picking a renderer from BFL’s internal ranking |
VEED still sells Image to Video at /tools/image-to-video-ai next to a separate script generator. Kapwing still sells Image To Video at /ai/image-to-video next to Script Generator. Canva’s public nav still splits AI image generator from AI video generator. BFL’s own site splits Klein from FLUX 3 Video. Your approval chain should look like those URL splits, not like one Comfy graph that “does everything.”
ClipCanva’s public FLUX 3 page on 19 October still says FLUX 3 is not currently connected. The image to video page tells you to check the current selector; the generator starts from one reference image. A 15 October open-weights drop on bfl.ai is not a model ID on clipcanva.com.
Creator / operator checklist
- Read the 15 October sentence out loud. Open weights. 5.6B. Text-to-image, edit, up to 10 still refs. Native 4K. Bounding boxes. If Slack says “FLUX 3 video, 4K Reel, ten keyframes, we beat Ideogram,” stop and split the sentence.
- Name the surface. BFL blog ≠ Klein model page ≠ full FLUX 3 Image page ≠ FLUX 3 Video ≠ BFL API
flux-3-image-klein≠ Hugging Face base vs distilled ≠ ClipCanva selector. Screenshot the picker you will actually use. - Cap the line before you draw boxes. Use the AI script generator. One sentence. Strike Elo claims, competitor names, and any “4K HDR commercial” Finance did not sign. A bounding box will not know which four words Legal locked.
- Lock frame one if the SKU must hold. Approved still, then image to video. Ten style refs are not the pack shot unless one of them is the pack shot.
- Write four beats, then decide whether you even need ten stills. Steal language from prompt ideas. For an 8-second Short you need four cards, not a 10-image collage. Klein’s boxes place type on a still. They do not time the end card.
- Count references like a budget. BFL’s cap is 10 images. Empty slots are fine. Ten mediocre refs are worse than one signed pack.
- Do not treat distilled 4-second 1MP as the edit. That NVIDIA ComfyUI number is still-image latency. Pixel restoration “preserves unedited regions” on the API still. Neither is motion review.
- Do not fine-tune, then skip Legal. Base weights, LoRA recipes, and “under $5M Kommunity” are rights and infra. They are not a signed line.
- Do not promise ClipCanva has Klein. The live FLUX 3 page says it is not currently connected. Check the selector. A competitor launch is not your dropdown.
- Compare the renderer after the still exists. Use Compare and generate on image to video or the AI video generator. A 4K PNG is a first frame option. It is not the brief.
Worked example: a Friday Short that does not start in a 10-ref zip
Say the job is an 8-second product Short. Legal line approved: Two-pack is back Friday. Same recipe. No new flavor. Someone opened the Klein blog, saw “up to 10 reference images,” saw bounding boxes on a postage-stamp demo, saw “native 4K,” saw $0.009 per 1K, and told the channel “dump the brand zip into Klein, we have 4K now, animate it.”
That sentence mixes four desks.
The line is a writer object. Put the approved sentence on the AI script generator. Klein will letterform a box. It will not store that Legal signed this sentence and struck the mango variant. Do not typeset “beats Ideogram V4.5” because BFL’s internal table did.
The still is a first-frame object. If the label must stay readable, generate or edit one pack shot — base or distilled, API or Comfy, your call — then put that file as frame one on image to video. Nine lifestyle refs and a three-view character sheet will not protect the nutrition panel. ClipCanva’s image-to-video page accepts one reference image. That is a product fact, not a dare to stitch ten.
The beat list is a card object. Steal shot language from prompt ideas, then write four lines:
- 0s — locked product still, no extra SKU
- 3s — one motion: pack rotates or the camera pushes; type stays readable
- 6s — end card with the signed sentence only
- 8s — stop. No orbit, no second flavor, no “4K open weights” lower-third
The 4K native still is a resolution object. If you generate 4K, you still crop to the Short. You do not owe the model a poster because the launch post is proud of 4K. If you later need layout control on the poster itself, then bounding boxes are the layout desk — on the still, with the same locked pack, not a new moodboard.
If the live ClipCanva selector has no Klein row, say so. Generate motion on image to video or the AI video generator. Keep the FLUX identity page at FLUX 3 for what is actually connected. Do not invent a Klein row on a site that has not listed one.
FAQ
Did FLUX 3 Image Klein actually ship, or is it still a teaser?
Black Forest Labs’ 15 October 2026 blog says “today we are releasing” Klein as open weights, with API access as flux-3-image-klein, Hugging Face base and distilled checkpoints, and a public model page. That is a still-image launch, not a video launch.
Is Klein a video model? Not on the pages cited here. The 15 October post describes text-to-image, image editing, and up to 10 reference images. BFL keeps a separate FLUX 3 Video URL. Do not brief Klein as a clip generator.
Can I put “beats Ideogram V4.5” in the ad because BFL ranked Klein above it? No. That ranking is BFL’s internal evaluation, on still-image tasks, with their stated caveats (parameter counts exclude text encoders; Krea 2 skipped for editing). It is not a claim Legal signed.
Does “pixel-perfect” or pixel restoration mean the pack shot cannot warp in video? No. The API note is that pixel restoration “preserves unedited regions of an image exactly.” Image-to-video is a different desk. Re-inspect the label after you animate.
Does the Kommunity License mean we can ship client work from local weights? Only inside what BFL published: free commercial use for individuals and companies with annual revenue under $5M, plus research and non-commercial use. Larger organizations need BFL’s self-hosted licensing or the API. Read the license. Do not take this paragraph as counsel.
Does Klein on Hugging Face mean ClipCanva has FLUX 3 Image Klein? No. ClipCanva’s public FLUX 3 page on 19 October still says the model is not currently connected. Check the live selector on image to video. A weights drop is not your dropdown.