Tasks and inputs
FLUX 3 Image Klein is listed for Image-text-to-image, Image-to-image, Text → Image. It accepts text, image and produces image. Use the publisher’s input format and preprocessing for your chosen deployment.
SyncDev model directory · Profile updated Nov 17, 2026
FLUX 3 Image Klein is BFL's open 5.6B editing-first image model: runs on an RTX 3090, ten references, bounding boxes, from $0.008 per image.
The specification score summarizes recorded features and limits. It is not a benchmark result.
Pricing conditions: API per image, generation and editing alike: 768sq $0.008, 1k $0.009, 1.5k $0.011, 2k $0.014, 4k $0.036. Weights free under the FLUX Kommunity License. Indicative; confirm on your BFL account.
Recorded modality support shows what this model can accept and produce. An undocumented modality is shown as unknown rather than assumed to be unsupported.
Accepts
Produces
Under a cent per image on the API, or free on your own GPU. FLUX 3 Image Klein is the compact, open-weights member of the FLUX 3 family, released by Black Forest Labs on 15 October 2026. It has 5.6B parameters, it runs on consumer cards down to an RTX 3090, and although it is built mainly for editing, it also generates from text, takes up to ten reference images and honours bounding-box layouts.
| Klein | Klein Base | |
|---|---|---|
| Training | Step-distilled | Undistilled |
| Steps / CFG | 16 / 1 | 50 / 3.5 |
| Purpose | Fast inference | Fine-tuning, LoRA, research |
| Where | API and Hugging Face | Hugging Face only |
| Max output | 4K | 4K |
A ComfyUI-ready repository (flux-3-image-klein-comfy) sits alongside both. The licence is the FLUX Kommunity License, not Apache; read it before commercial self-hosting.
Klein was trained on long scene captions produced by a prompt upsampler, so BFL's advice is blunt: always run it. The API and the Diffusers modular pipeline do so automatically, and the CLI does too unless you pass --no-pup. Running locally, remember the upsampler is a separate 9B model and needs its own GPU memory, which matters on a 24GB card.
Klein speaks the FLUX 3 Image request format. Change the path to /v1/flux-3-image-klein and an existing request runs on the smaller model. It adds seed, steps and output_format. It lacks match_input; edits take the size of the first reference instead, and width and height start at 64 px. Generation and editing cost the same:
At 4k that is roughly a seventeenth of the full model's price. Prices are indicative.
Train on Klein Base, then load the result into distilled Klein without conversion. BFL's reference recipe: rank 32, alpha 32, AdamW, learning rate 1e-4 with 200 warm-up steps, 4,000 steps on about 20 images at roughly 1MP with captions ending in a trigger word. That took about 4.5 hours on one GPU. At inference use 16 steps, CFG 1 and a LoRA scale near 0.75. AI Toolkit supports it under the flux3 architecture.
Often, for edits. Run your twenty most common editing jobs (background swaps, colour changes, object removal) through both models and score them side by side. Then try your hardest multi-reference compositions; that is where the larger full FLUX 3 Image (see Black Forest Labs) and rivals such as GPT Image 2.5 Flare may still pull ahead. For local use, time a 2k edit on your actual card, upsampler included. More options are under image-to-image models.
The Hugging Face repositories went up on 13 October, two days before the official release.
Compare the model on the work you need to complete.
FLUX 3 Image Klein is listed for Image-text-to-image, Image-to-image, Text → Image. It accepts text, image and produces image. Use the publisher’s input format and preprocessing for your chosen deployment.
Choose the output size and quality your design needs. Include drafts and revisions when budgeting an accepted image.
The weights are downloadable under FLUX Kommunity License. Check runtime support and size the complete model and working context for your expected concurrency.
These records share a directory category. Check the source card, output type and deployment requirements before treating a model as a replacement.
ideogram
Ideogram 4.5 is an editing-focused image model that resists drift across repeated edits, with native 2K output, Precise Edit and fixed per-image pricing.
1038lab
Hugging Face repository for Image-to-image.
Alissonerdx
Alissonerdx/BFS-Best-Face-Swap is a source-linked Hugging Face repository assigned to Image-to-image. Its recorded task interface accepts image and produces image.
autoweeb
Hugging Face repository for Image-to-image.
black-forest-labs
black-forest-labs/FLUX.1-Kontext-dev is a source-linked Hugging Face repository assigned to Image-to-image. Its recorded task interface accepts image and produces image.
black-forest-labs
black-forest-labs/FLUX.2-dev is a source-linked Hugging Face repository assigned to Image-to-image. Its recorded task interface accepts image and produces image.