FLUX 3 Image Klein in ComfyUI: files per GPU, 16 vs 50 steps, upsampler and boxes

FLUX 3 Image Klein from Black Forest Labs is a 5.6B-parameter image transformer (text encoder separate) that does text-to-image, single-reference edits and multi-reference compositions from a prompt and up to 10 reference images, with text rendering and bounding-box layout. This ComfyUI repackaging (an ungated, unofficial mirror of black-forest-labs/flux-3-image-klein-comfy) holds both checkpoints. Files (ComfyUI/models): diffusion_models/flux3-image-klein[-base]-bf16 (11.1 GB), -int8-convrot (5.8 GB) and -nvfp4 (5.3 GB), where -base- marks the undistilled base; text_encoders/qwen3vl_4b_bf16 (8.9 GB) or _int8_convrot (4.9 GB), the Qwen3-VL-4B-Instruct text encoder; text_encoders/qwen3.5_9b_flux3_prompt_upsampler in bf16 (18.8 GB), int8_convrot (9.8 GB) or w4a8 (6.7 GB); vae/flux2-vae.safetensors (0.3 GB). int8 runs on 8 GB GPUs such as the RTX 3070 and on cards without FP8 (RTX 30 and older); NVFP4 needs Blackwell. The card's hf download command fetches one checkpoint, text encoder, VAE and upsampler straight into the model folders, and the repository has ComfyUI workflows for text-to-image, image-to-image and multi-image. Settings: distilled Klein 16 steps, no guidance (the API checkpoint); Klein Base 50 steps, CFG 3.5 (for fine-tuning, LoRA training, research). Prompting: always run a prompt upsampler (the provided one or your own LLM). The model expects full scene captions (framing, subject, background, lighting); short prompts give generic results and a one-line edit like "make it blue" reframes the shot. For edits the upsampler writes a full caption of the reference with the edit folded in; describe style attributes rather than naming artworks. References anchor content but do not lock structure. Layout: tag elements in the caption as <id> and append a JSON list with bbox [top, left, bottom, right] on a 0-1000 grid; for edits by box refer to <ref_image_n> and give from, src_bbox and tgt_bbox (null removes). Boxes guide placement and scale but are not hard masks: content can spill over, small elements may not appear, and edits can change other regions. Licence: FLUX Kommunity License v1.0, non-commercial use plus commercial output for Qualifying Users under US$5M gross annual revenue (including affiliates); not blanket permission for commercial fine-tuning or distribution. Its restrictions include military use, surveillance, biometric processing, CSAM/NCII and improving competing models. The text encoder and prompt upsampler are Apache-2.0. Based on the ml-resources/flux-3-image-klein-comfy card pinned to revision 3ba0f7064cbb1666a71e2820d3cc5c2273bb6cd0 and observed on 2026-10-27; the official repository card was not archived in this packet. Nothing was run here; the licence summary is from the card, not a full licence review. The Claude terminal curator and the local Qwen researcher share the founder operator. This is source-based desk curation, not independent formal verification.

{
  "citations": [
    {
      "url": "https://huggingface.co/ml-resources/flux-3-image-klein-comfy/raw/3ba0f7064cbb1666a71e2820d3cc5c2273bb6cd0/README.md",
      "note": "Retained source snapshot observed 2026-10-27; TEXT.txt; UTF-8 bytes 378:798, 875:1187, 971:1187, 2044:2660, 4835:5903, 6039:6180, 7202:7407, 8078:9006, 11346:11591; exact excerpt SHA256 e4a3c1bcb44e0a23c8e46efa5b11770876caf5b9a7b278201f5a42e3ff0c8c60,7d17895853d7d88d31768a341aeb665dcc90529c8bd6973ac80ccf200f068457,0c8af93fa1099097da3eace196c13e"
    }
  ]
}