LAXMAYDAY/flux2-klein4b-sdr2hdr-logc4-lora

SDR → HDR LoRA for FLUX.2-klein-4B

Turn any SDR photo into a scene-referred HDR image (ARRI LogC4 encoding, linear values recoverable up to ~470x). Clipped skies, suns, lamps and speculars are reconstructed instead of staying flat white — ready for EXR export, grading, or IBL.

Highlights

• Prompt-controlled intensity: four control words dial how aggressively highlights are reconstructed, from subtle to sun-level.
• Faithful, not hallucinated: on held-out data the median per-pixel relative error is ~10%, global exposure is self-calibrated, and low-dynamic scenes stay low — the model does not invent highlights that were never there.
• Two formats included: sdr2hdr-klein4b-v1.safetensors (kohya/musubi naming) and sdr2hdr-klein4b-v1-COMFYUI.safetensors (ComfyUI-native keys).

Quick start (ComfyUI)

1. Put sdr2hdr-klein4b-v1-COMFYUI.safetensors in models/loras/, with FLUX.2 klein-4B as the base model.
2. Load it with a standard LoRA loader (strength 1.0) and feed the SDR photo as the reference/edit image.
3. Prompt: convert to scene-linear HDR, LogC4 encoded, intense highlights (swap the last words per the table below).

Important — reference image size: resize (aspect-fit + center-crop) the SDR input to exactly the generation resolution before conditioning. A larger reference makes the model reproduce content at 1:1 scale and the output turns into a zoomed crop.

Intensity control

control word	reconstruction target
extreme peak luminance	sun-level sources, up to the LogC4 ceiling (~470x)
intense highlights	strong practical lights, tens of x
strong highlights	pronounced but moderate highlights (~5x)
soft dynamic range	gentle expansion for low-contrast scenes (~2-3x)

Fixed prompt template: convert to scene-linear HDR, LogC4 encoded, {control word}

Output format

The generated PNG stores raw ARRI LogC4 code values, not sRGB (it looks flat grey in a normal viewer — that is correct). Decode to scene-linear:

V = png / 255
L = (2^(14*(V - c)/b + 6) - 64) / a
a = 2231.8259, b = 0.9071358748778103, c = 0.0928641251221897

Mid-grey maps to 0.18; values above 1.0 are HDR highlights. Write the linear buffer to a 32-bit EXR for VFX / grading / IBL use.

musubi-tuner inference

python src/musubi_tuner/flux_2_generate_image.py \
  --model_version klein-4b \
  --dit flux-2-klein-4b.safetensors --vae ae.safetensors \
  --text_encoder text_encoder/model-00001-of-00002.safetensors \
  --control_image_path input_sdr_fitted.png \
  --prompt "convert to scene-linear HDR, LogC4 encoded, intense highlights" \
  --image_size 768 1360 --infer_steps 28 \
  --lora_weight sdr2hdr-klein4b-v1.safetensors \
  --save_path out --output_type images

Training & evaluation

LoRA (dim 32) trained with musubi-tuner on ~1k human-curated SDR/HDR pairs spanning natural scenes, graded cinematic footage and film-production plates. On held-out samples: median in-band per-pixel relative error ~0.10, exposure gain 0.87–1.14 (no post-alignment needed), and monotonic peak response across all four intensity words. See samples/ for input / output / ground-truth comparisons (shown at −3EV to reveal highlight structure).

Limitations

• Slight highlight overshoot above ~100x on extreme content.
• Trained against synthetic tonemaps; heavily stylized real-world SDR grades may behave as out-of-distribution input.
• Native resolution ~1MP; not intended for radiometric measurement.

License & access

Gated for non-commercial research. Training data has mixed provenance; verify upstream terms for your use case. Base model: FLUX.2-klein-4B (Black Forest Labs).