Why Klein needs full captions
FLUX 3 Image Klein is Black Forest Labs' 5.6B open-weights image model (released October 15, 2026): text-to-image, editing with up to 10 reference images, and bounding-box layout control in a single model, up to 4K. It was trained on the long captions produced by its own prompt upsampler — a 9B vision-language model fine-tuned from Qwen3.5-9B. In the official repository, BFL is blunt about it: do not use the model without the upsampler, because a prompt sent as typed gives noticeably weaker results.
That leaves two ways to get good prompts. Run the official pipeline and let the 9B upsampler rewrite your idea (it needs GPU room next to the model), or write the upsampled caption yourself and pass it with --no-pup. This builder takes the second route: it assembles the same kind of caption the upsampler writes — medium, shot, subject detail, setting, light, palette — from a short form, in your browser, with no GPU and no account.
- Caption builder — short form in, 100+ word upsampler-style caption + ready CLI command out.
- Layout boxes — draw boxes on a canvas, export the official
boxes.json format.
- Edit prompt — structured final-frame descriptions for edits with up to 10 references.
Distilled vs Base
| FLUX 3 Image Klein (distilled) | FLUX 3 Image Klein Base |
|---|
| Steps / guidance | 16 steps, no guidance | 50 steps, CFG 4 |
| Speed | ~6× faster, the API checkpoint | Slower, reference quality |
| Best for | Generation & editing, self-hosting | Fine-tuning, LoRA training, research |
| Extras | Fixed settings, nothing to tune | Negative prompts, full control |
Both share the same weights family, text encoder and bounding-box support. Quantized FP8 / NVFP4 builds ship with the release, and the ComfyUI int8 files run on 8 GB GPUs such as an RTX 3070.
The bounding box format
Layout control uses a boxes.json file: a list of boxes in [y_min, x_min, y_max, x_max] order, coordinates from 0 to 1, origin at the top-left. Each box has an id the caption can reference and a desc of what goes inside. This is the official example from the BFL repository:
boxes.json — official example
[{"id": "obj_1", "desc": "a red hot-air balloon", "bbox": [0.08, 0.10, 0.45, 0.42]},
{"id": "obj_2", "desc": "a lighthouse on a cliff", "bbox": [0.30, 0.60, 0.95, 0.90]}]
Boxes guide placement and scale rather than acting as hard pixel masks — an element can spill slightly past its box. For edits, a box can also carry "from": "ref_image_0" (or "target") with src_bbox and tgt_bbox to move an element from a reference into a new position. The Layout tab above exports exactly this format.
Frequently asked questions
What is FLUX 3 Image Klein?
Black Forest Labs' 5.6B-parameter open-weights image model, released October 15, 2026. One model does text-to-image, editing with up to 10 reference images, and layout control with bounding boxes, at resolutions up to 4K.
Why do my Klein prompts look generic?
Klein is trained on long captions written by its prompt upsampler. A short prompt sent as typed gives noticeably weaker results — BFL says not to use the model without the upsampler. Either let the official CLI upsample, or write a full caption yourself (this builder assembles one) and pass it with --no-pup.
What is the bounding box format?
A boxes.json list of {"id", "desc", "bbox": [y_min, x_min, y_max, x_max]} with 0–1 coordinates and a top-left origin. Edit boxes can add "from", "src_bbox" and "tgt_bbox". The Layout tab exports this format directly.
Distilled or Base — which should I use?
Distilled for generation and editing (16 steps, no guidance, ~6× faster). Base (50 steps, CFG 4) for fine-tuning, LoRA training and research.
Can Klein edit images?
Yes — up to 10 reference images, and the first reference sets the output size. Describe the final frame rather than only the change, and state what must stay unchanged.
Does this tool generate images?
No. It builds captions, boxes.json and CLI commands in your browser. You run the model yourself via the official CLI, ComfyUI, Diffusers or the BFL API.