You can deploy this model and start earning money today!

cusiman/
Boogu-INT8

Boogu/Boogu-Image-0.1-Turbo

finetune

apache-2.0

Views

No views yet

Boogu-Image-0.1-Turbo-hotfix — INT8 Quantized

INT8 tensor-wise quantization of Boogu-Image-0.1-Turbo-hotfix, produced with convert_to_quant.

This is the four-step distilled Turbo variant of the Boogu-Image-0.1 family, quantized to INT8 for reduced VRAM usage and faster loading while preserving output quality.

Quantization Details

Format: INT8, tensor-wise scaling (int8_tensorwise)
Method: Simple quantization (no learned rounding / SVD optimization)
ConvRot: Not applied — Boogu's attention/feed-forward layer dimensions (840, 3360) are not compatible with ConvRot's group-size requirements (group size must be a power of 4 and evenly divide in_features)
Metadata: Includes comfy_quant metadata for native ComfyUI compatibility

Excluded Layers (kept in BF16)

The following layers were kept at full precision rather than quantized, based on community guidance for this architecture:

image_index_embedding
ref_image_patch_embedder.weight
*.norm1.linear.weight (all blocks)
norm_out.linear_1.weight
norm_out.linear_2.weight

These are embedding and normalization/modulation layers, which are commonly excluded from quantization to preserve generation quality and avoid instability.

Conversion Command

bash

ctq -i boogu_image_turbo_hotfix_bf16.safetensors \
    -o boogu_image_turbo_hotfix_int8.safetensors \
    --int8 --scaling_mode tensor --simple --low-memory \
    --comfy_quant --save-quant-metadata \
    --exclude-layers "(image_index_embedding|ref_image_patch_embedder|norm1\.linear|norm_out)"

Usage in ComfyUI

Place the .safetensors file in ComfyUI/models/diffusion_models/
Load it with the Load Diffusion Model (UNETLoader) node
Use the standard Boogu Turbo workflow:
- Text encoder: qwen3vl_8b_fp8_scaled.safetensors
- VAE: flux1_vae_bf16.safetensors
- Steps: 4 (Turbo default)
Note: The first generation after loading may take significantly longer (several minutes) due to one-time kernel warmup for this model's tensor shapes. Subsequent generations run at normal speed.

Requires a reasonably recent ComfyUI build with INT8 tensor-wise (int8_tensorwise) support. If you hit a KeyError related to quantization format on load, update ComfyUI or check that your build supports this format.

Credits

Base model: Boogu-Image-0.1 by the Boogu team (Apache-2.0)
Quantization tooling: silveroxides/convert_to_quant

License

Apache-2.0, inherited from the base Boogu-Image-0.1 model.

cusiman/Boogu-INT8