PRACTICAL GUIDE · PROMPTTOVISUAL
FLUX 3 Image & Klein: ComfyUI Workflows, Weights and API
Download an official FLUX 3 Klein workflow, copy its settings and file paths, then follow the steps in your own ComfyUI. Includes model, GPU and license requirements.
TAKE THE WORKFLOW · RUN IN YOUR OWN COMFYUI
Download a workflow. Prepare your setup.
Get the official workflow and its matching settings below. You need your own compatible ComfyUI installation and model files; this page does not generate images.
JSON only · under 100 KB at review · no model weights · no sign-in or PTV credits. On a phone, save the file and transfer it to your ComfyUI computer.
Open official JSON · Inspect the official template
View or manually copy the selected settings
2. Put the four model files in place
Update to ComfyUI v0.41.3 or a later compatible version, including its template dependencies. Update instructions. Obtain weights from BFL after reviewing its access and license terms; these are separate, large downloads.
- Image model
- flux3-image-klein-int8-convrot.safetensorsPlace in
ComfyUI/models/diffusion_models/ - Text encoder
- qwen3vl_4b_int8_convrot.safetensorsPlace in
ComfyUI/models/text_encoders/ - Prompt upsampler
- qwen3.5_9b_flux3_prompt_upsampler_w4a8.safetensorsPlace in
ComfyUI/models/text_encoders/ - VAE
- flux2-vae.safetensorsPlace in
ComfyUI/models/vae/
These are the INT8 / W4A8 template selections. The 8 GB statement below is BFL’s hardware claim, not a guarantee for your resolution or setup. Model downloads and local or rented compute have their own requirements.
3. Load, run and find the result
- Drag the downloaded JSON onto your ComfyUI canvas. If a model is missing, check its filename and folder in the list above, then refresh model lists or restart.
- Check the selected checkpoint, 16 steps, No CFG guidance, Euler/simple and 1024 × 1024. Keep prompt upsampling on. Use the exposed workflow controls, not an unrelated sampler inside a subgraph.
- Replace the template’s prompt with your scene description. The starter below is an editable example; we have not generated or benchmarked it.
- Choose Run in ComfyUI. After a successful run, inspect the Save Image node and your ComfyUI output folder. If it fails, read the node error and check missing files or GPU memory before another run.
Read or edit the starter prompt in your own workflow
This guide checks the official files and helps you prepare a local run. It does not certify an image result, GPU benchmark or successful model execution.
FLUX 3 Image Klein is a separate open-weight release
BFL released the 5.6B Klein image transformer on October 15, 2026, in base and distilled variants. Its separate text encoder and optional prompt upsampler add to the hardware requirement. Klein weights do not establish open-weight availability for the larger FLUX 3 Image model covered below. BFL advertises multi-reference editing and native 4K for Klein; these are vendor capabilities, not results measured here.
Why use the corrected ComfyUI v0.41.3 workflows?
The ComfyUI team says it reworked the launch examples with BFL for v0.41.3: prompt upsampling is enabled by default, base text-to-image changes from 25 to 50 steps, and distilled from 15 to 16. Editing examples also gained prompt upsampling. Update the engine and template dependencies, then load a fresh Klein template; an older saved graph will not acquire the new settings automatically. We inspected the official JSON files, but did not install weights or run inference.
1. Update ComfyUI, including its templates
Use v0.41.3 or a later compatible release. On Windows Portable, the documented update/update_comfyui_stable.bat selects stable releases; verify the resulting version. On Desktop, update the engine through Manage → Update; the desktop application version is separate. For a manual installation, activate its Python environment, run git pull, then pip install -r requirements.txt and restart. Updating code alone can leave old workflow templates installed.
2. Match the distilled template to its checkpoint
The official image_flux3_klein_distilled_t2i.json selects flux3-image-klein-int8-convrot.safetensors, 16 steps, Euler and the simple scheduler at 1024 × 1024, with prompt upsampling enabled. Its text encoder is qwen3vl_4b_int8_convrot.safetensors; its VAE is flux2-vae.safetensors; its upsampler is qwen3.5_9b_flux3_prompt_upsampler_w4a8.safetensors. These are inspected template defaults, not our tested optimum.
3. Use the base workflow for the base checkpoint
The separate image_flux3_klein_t2i.json selects flux3-image-klein-base-int8-convrot.safetensors, 50 steps and CFG 3.5, with the same encoder, VAE and W4A8 upsampler. Do not feed base weights into the distilled settings or rename one variant to masquerade as another. Check the exposed subgraph controls: an internal sampler default alone does not describe the effective workflow settings.
4. Install the matching files in the correct folders
Place the Klein checkpoint in ComfyUI/models/diffusion_models, both Qwen components in models/text_encoders, and flux2-vae.safetensors in models/vae. Use the official repository and its access/license process. Restart or refresh model lists, then select the matching files. Base and distilled offer BF16, INT8 and NVFP4 variants.
Will Klein fit on an 8 GB GPU?
BFL reports INT8 operation on an RTX 3070 with 8 GB VRAM. This does not guarantee every resolution, reference count or upsampler configuration. RTX 30-series lacks FP8 support; NVFP4 targets Blackwell. Download size is not peak VRAM. We have not measured memory, speed or 4K operation.
Can Klein be used commercially? Read the exact license
The image weights use the FLUX Kommunity License, not an unrestricted Apache grant. Its commercial output-generation permission applies to qualifying entities with total annual revenue below USD 5 million, including affiliates; other uses and larger entities require checking the applicable terms or a separate license. The license also requires content filtering or output review before distribution, with applicable AI disclosures. The separate Qwen components do not determine the image-weight license.
Why is the result poor, or a loader missing?
Compare checkpoint variant, steps, upsampling and component selections with the official template. Missing templates warrant checking dependencies; missing model entries warrant checking folders and filenames. Without upsampling, BFL expects a full-scene caption, not just a short edit instruction. These are documentation-based checks, not findings from our own inference run.
The larger FLUX 3 Image: can I use it here?
For generation in PromptToVisual, open our FLUX 3 Image model page. The studio supports square 1024 × 1024 text-to-image with the full model, at 5 credits per image. Reference editing and 4K are not available here. The Klein workflow above remains a separate setup for your own ComfyUI; downloading that JSON does not generate an image in PromptToVisual.
Use up to ten references with explicit roles
BFL supports up to ten input images. Its model page assigns reference tokens in upload order, beginning with ref_image_0. More references are a control option, not proof that every identity or logo survives.
Bounding boxes control layout, not a hard clipping boundary
The official box format uses [top, left, bottom, right] coordinates on a 0–1000 grid. Reference tokens and source/target boxes can describe where an element starts and where it should appear. The structured rows go inside prompt; they are not a separate box parameter. BFL explicitly says boxes guide placement and scale rather than act as clipping masks. Keep the aspect ratio used to design the layout.
What does pixel-perfect editing mean here?
Pixel-perfect is BFL’s product claim, not our guarantee for every edit. Its editing documentation says pixels outside boxes usually remain unchanged, while shadows, reflections or nearby lighting may change. Inspect these surrounding effects as well as the selected object. Do not equate a bounding box with Ideogram’s explicit black-and-white mask mechanism.
Native 2K and 4K: resolution class versus exact dimensions
BFL advertises native 2K and 4K rendering. These labels do not imply one fixed width for every aspect ratio; its 4K example is 5456 × 3072.
What does the official API accept and return?
The documented endpoint is POST https://api.bfl.ai/v1/flux-3-image, authenticated with x-key. Input references may be URLs or base64; the overview lists 256 × 256 through 16 MP per image. It lists resolution options 768sq, 1k, 1.5k, 2k and 4k. Poll the returned polling_url until Ready and save result.sample within its documented one-hour window. This describes BFL’s API, not an integration implemented here.
Larger-model API price snapshot: October 2, 2026
At the October 2 check, the official table listed USD 0.041 per image at 768sq, 0.048 at 1k, 0.100 at 2k and 0.607 at 4k. The table reviewed does not give a 1.5k price, although that option appears in the endpoint overview: do not interpolate it. BFL says the submission response includes request cost. These are provider/API prices, not PromptToVisual retail prices. These prices were not refreshed in this Klein workflow update; recheck the current table before spending. They are not Klein prices.
Best fit by task: FLUX 3 Image vs Ideogram 4.5
For multi-reference layouts, assess FLUX’s documented spatial controls. For a targeted mask and iterative local revision, inspect Ideogram’s distinct Precise Edit contract. Multi-turn drift remains untested here. If deployment rights matter, request the exact license and delivery terms rather than selecting on the phrase open weights. Neither model is declared best overall.
How to review a layout without inventing a benchmark
Prepare a short brief with required objects, rough positions, lettering and output dimensions. Keep each source and identify which details must not change. A review can inspect count, placement and identity separately. That checklist is an editorial suggestion; it is not evidence that either model passed it.
Choose a workflow, then read its contract
These are source-based guides. PromptToVisual alternatives use other listed models, not Ideogram 4.5 or FLUX 3 Image.
Ideogram 4.5 precise editing · FLUX 3 Image composition · PTV image editor alternative · PTV image combiner alternative · Product photography · Retained color-edit examples · Retained composition examples · All guides
Sources and next steps
Klein workflows require compatible model files and sufficient local hardware. Review the applicable model license; the larger-model API has separate terms.
Klein release and model scope →
Why use the corrected ComfyUI v0.41.3 workflows? →
1. Update ComfyUI, including its templates →
2. Match the distilled template to its checkpoint →
3. Use the base workflow for the base checkpoint →
4. Install the matching files in the correct folders →
Will Klein fit on an 8 GB GPU? →
Can Klein be used commercially? Read the exact license →
Why is the result poor, or a loader missing? →
Flare generator and real samples → · Sunburst generator and real samples → · Images 2.5 and Sketch · LLaDA ComfyUI setup
Continue with the right tool
AI image combiner · Stitch images · Photo collage · Nano Banana 2.1 guide · Related composition guide