Stick9190/kroma-v0.2-turbo-int8-convrot-learned

Kroma v0.2 Turbo - INT8 ConvRot

Learned-rounding INT8 ConvRot quantization of Kroma v0.2 Turbo, a Krea.2-based diffusion model by Lodestones. The 25.64 GB BF16 checkpoint is reduced to approximately 13.49 GB.

Conversion

• Tool: silveroxides/convert_to_quant
• Format: INT8 row-wise W8A8 with embedded ConvRot metadata
• ConvRot group size: 256
• Method: learned rounding (AdaRound) with low-memory streaming conversion
• Quantized: 224 attention and MLP matrices across all 28 transformer blocks (~12.16B parameters, or 94.81% of the model)

The image input and final output projections, timestep embedding and modulation paths, text projection, and text-fusion transformer remain BF16. Norms and biases also retain their original precision. This protects the shared conditioning and boundary paths while quantizing the large transformer core.

Command

ctq -i <input-model>.safetensors -o kroma-v0.2-turbo_int8_convrot_learned.safetensors `
  --krea2 `
  --int8 --scaling_mode row `
  --convrot --convrot-group-size 256 `
  --comfy_quant --save-quant-metadata `
  --low-memory --device cuda

Quantization is lossy, so outputs are not bit-identical to the original BF16 checkpoint.