777Radik/Qwen-Image-2512-FP8-E4M3FN

Qwen-Image-2512 FP8 E4M3FN transformer

Transformer-only FP8 storage conversion of Qwen/Qwen-Image-2512. Floating tensors are stored as float8_e4m3fn; shard names, tensor keys, index, config, and non-floating tensors are preserved for DiffSynth-compatible loading.

Files

• 9 safetensors shards
• diffusion_pytorch_model.safetensors.index.json
• config.json
• quantization-manifest.json with source and output SHA-256 for every shard

The transformer is 20,430,631,776 bytes, down from 40,861,027,880 bytes (50.0% smaller). The base tokenizer, text encoder, and VAE are not duplicated; load them from Qwen/Qwen-Image-2512.

Validation

The same 100-sample Russian held-out benchmark was generated with the source BF16-storage transformer and this FP8-storage transformer. Both runs used the same trained ControlNet, prompts, control images, 20 steps, and per-sample seeds.

• 100/100 generated PNG SHA-256 hashes matched
• manual exact Cyrillic score: 99/100 before and after conversion
• regression: 0 percentage points
• OCR diagnostic exact: 77/100 before and after conversion
• mean OCR CER: 0.1142 before and after conversion
• mean glyph similarity: 0.9564 before and after conversion

The remaining manual failure was ЕЩЁ → ЁЩЁ; it is present before quantization and is not a quantization regression.

Scope

This is a storage conversion, not a new base model and not an SDNQ graph rewrite. SDNQ was not used because the selected DiffSynth Blockwise ControlNet pipeline expects the original Qwen transformer graph and keys.

Use at your own risk outside the tested DiffSynth pipeline and Blackwell FP8 environment.