Qwen-Image-2512 Realism (bf16, LoRA-merged UNet)

A single-file bf16 diffusion model (UNet only) for ComfyUI: Qwen/Qwen-Image-2512 with two photorealism LoRAs baked into the weights. No LoRA loaders needed at inference — load it like any other diffusion model and you get the realism the two LoRAs provide.

The main reason this exists: it is one file to pull onto a fresh training/inference box instead of "base model + LoRA A + LoRA B + remembering the strengths".

Merge recipe

Merged in ComfyUI (UNETLoader → LoraLoaderModelOnly → LoraLoaderModelOnly → ModelSave):

The full merge graph is embedded in the .safetensors header (__metadata__.prompt / __metadata__.workflow), so the recipe is verifiable from the file itself:

python -c "import json,struct;f=open('qwen-image-2512-realism-bf16.safetensors','rb');n=struct.unpack('<Q',f.read(8))[0];print(json.loads(f.read(n))['__metadata__']['prompt'])"

Files

Tensor keys are unprefixed (img_in.weight, transformer_blocks.0.…) — the model.diffusion_model. prefix that ComfyUI's ModelSave writes has been stripped. ComfyUI loads either form; some trainers (e.g. ostris/ai-toolkit) expect the bare keys, so this is the more portable of the two.

INT8 ConvRot

qwen-image-2512-realism-int8_convrot.safetensors is an INT8 ConvRot conversion of qwen-image-2512-realism-bf16.safetensors for ComfyUI. It loads with the same Load Diffusion Model (UNETLoader) node (weight_dtype default) and needs a ComfyUI with comfy-kitchen INT8 ConvRot support.

Conversion recipe

The reference is supermind/int8_convrot_models diffusion_models/qwen/qwen_image_2512_int8_convrot.safetensors (sha256 cd6f8c0388cdf88dd10f5b5516028cd39ee59bddc8fed402d391d7cb29169ee2). Its header defines the per-key rule (which tensors are quantized, the .comfy_quant JSON, dtypes), mirrored key-for-key.

INT8 ConvRot (840 tensors): int8_tensorwise with a per-row F32 weight_scale of shape [N, 1] and ConvRot enabled.

Not quantized (1093 tensors): img_in, txt_in, proj_out, norm_out.linear, time_text_embed.timestep_embedder.linear_1/2, all biases and all norms — BF16, bit-identical to the bf16 file. These are exactly the tensors the reference keeps unquantized.

Each quantized layer carries a .comfy_quant U8 tensor with the same JSON as the reference: {"format": "int8_tensorwise", "convrot": true, "convrot_groupsize": 256}.

Comfy-Org's own INT8 ConvRot build of the same architecture (qwen_image_edit_2511_int8_convrot.safetensors in Comfy-Org/Qwen-Image-Edit_ComfyUI) uses the identical layout: same 840 quantized tensors, same group size, same dtypes. The only difference is an empty __index_timestep_zero__ marker tensor specific to the edit model, which is not added here.

Scale rule. The reference scales are not plain absmax: a per-row absmax quantizer (comfy-kitchen's TensorWiseINT8Layout.quantize) changes 37–73 % of the reference's int8 values on the layers checked. The rule was recovered from the reference itself and reproduces it (see Verification):

1. Upcast the BF16 weight to FP32 and apply the ConvRot rotation W · Hᵀ (block Hadamard, group size 256) with comfy-kitchen's portable implementation.
2. For every output row r, try the 80 clip ratios c = torch.linspace(0.55, 1.0, 80), with s = absmax(r) · c / 127 and q = clamp(round(r / s), −127, 127).
3. Keep the c with the smallest Σ (q·s − r)²; on an exact tie the smaller c wins. Store q as I8 and s as the F32 weight_scale.

comfy-kitchen 0.2.35, CUDA device. __metadata__ records the source file name, source sha256, recipe and reference.

Verification

Reproduction of the reference. The same script, run on the vanilla bf16 base (qwen_image_2512_bf16.safetensors from Comfy-Org/Qwen-Image_ComfyUI), was compared with the supermind INT8 file tensor by tensor:

The one row with a different clip ratio is an FP32 near-tie: the two candidate ratios give quantization errors that differ by 1.5 × 10⁻⁷ relative, about one ulp of the sum, so the choice depends on summation order. The 1-ulp scale differences have the same origin. Both are at the level of floating-point rounding, not a different recipe.

Dequantization check. Each INT8 tensor of this file was dequantized through ComfyUI's own comfy_kitchen path (inverse Hadamard rotation included) and compared with qwen-image-2512-realism-bf16.safetensors:

Layout check. 3613 tensors; key set, dtypes, shapes and all 840 .comfy_quant payloads are identical to the reference. Only __metadata__ differs.

Load check. comfy.sd.load_diffusion_model (the UNETLoader path, ComfyUI 1568e6c) detects QwenImage, 840 of 846 Linear layers loaded as int8_tensorwise (ConvRot, group size 256), no missing or left-over keys.

Usage (ComfyUI)

Put the file in ComfyUI/models/diffusion_models/ and load it with Load Diffusion Model (UNETLoader). It is a UNet only — you still need the stock Qwen-Image text encoder and VAE:

• Text encoder: qwen_2.5_vl_7b.safetensors → CLIPLoader, type qwen_image
• VAE: qwen_image_vae.safetensors → VAELoader

Anything that works with vanilla Qwen-Image-2512 (Lightning LoRAs, ControlNets, samplers) works here too.

Settings used for the example images

Two-pass, 4-step Lightning:

• ModelSamplingAuraFlow shift 3.1
• Qwen-Image-2512-Lightning-4steps-V1.0-bf16 LoRA @ 1.0
• Pass 1: 1024×1536, res_2m / bong_tangent, 16 steps, cfg 2.7, eta 0.5, denoise 1.0 (RES4LYF ClownsharKSampler)
• Latent upscale ×1.25 → 1280×1920, Pass 2: same sampler, denoise 0.45
• Negative (Chinese, from the Qwen-Image default set): 低分辨率，低画质，肢体畸形，手指畸形，画面过饱和，蜡像感，人脸无细节，过度光滑，画面具有AI感，构图混乱，背景模糊、散景。

A plain KSampler with the Lightning LoRA (4–8 steps, cfg 1.0) also works fine; the RES4LYF setup is just what the samples happen to use.

Examples

examples/merged/ — this model. examples/base/ — vanilla Qwen-Image-2512 for comparison, same seed (927056266523880), same scene, same sampler settings.

The example PNGs do not carry generation metadata.

Download

pip install -U huggingface_hub
hf download beycanai/qwen-image-2512-realism qwen-image-2512-realism-bf16.safetensors \
  --local-dir ComfyUI/models/diffusion_models

INT8 ConvRot:

hf download beycanai/qwen-image-2512-realism qwen-image-2512-realism-int8_convrot.safetensors \
  --local-dir ComfyUI/models/diffusion_models

Notes

• The bf16 file is not quantized. For less VRAM use the INT8 ConvRot file above, or quantize it yourself (GGUF / fp8) — nothing here blocks that.
• Merged LoRA strengths are fixed. If you want a different balance, merge again from the base with your own strengths rather than trying to unbake these.
• Useful as a LoRA-training base when you want the realism prior already present; it is a full-weight model, so it trains like the base does.

Credits

• Base model: Qwen/Qwen-Image-2512 (Apache-2.0)
• SamsungCam UltraReal and NiceGirls UltraReal — credit to their authors; their Civitai license terms apply to this merge as well.