Iris-3B INT4 ConvRot
Repackaged model files for ComfyUI and others, with INT4 ConvRot converted for testing in Radiant Canvas for fast macOS inference: Radiant Canvas on the App Store.
The original model is speridlabs/iris-3b, a 3B pixel-space diffusion transformer (no VAE) by Sperid Labs. See the original repository and the code for usage.
Files
iris-3b-int4_convrot.safetensors: the text-to-image diffusion transformer in ComfyUI's native ConvRot W4A4 tensor format. Tensor names are unchanged from the original checkpoint.config.yaml: the original architecture and sampling config.
The Qwen3-VL-4B-Instruct text encoder is not included. Get it from Qwen/Qwen3-VL-4B-Instruct. The depth and upscaler fine-tunes are not converted here.
At the time of conversion ComfyUI has no native Iris-3B model support, so this file needs a loader that knows the Iris architecture and the ConvRot W4A4 format.
Conversion details
- Original model revision:
7445443349bc9abe3c96f01ff793e2098ca012b3 - Source format: FP32 EMA weights, 639 tensors
- Quantization: comfy-kitchen 0.2.37
TensorCoreConvRotW4A4Layout, matching theint4_convrotpath of the Starnodes model converter - Quantized layers: 168 attention and MLP weights (image side of the 8 dual-stream blocks, and all 16 single-stream blocks)
- ConvRot group size: 256
- Quantization group size: 64
- Output size: 3,893,372,344 bytes
- SHA-256:
3e517eb8c08268409ce94d252897726a706b46471bdfa8284c2771ec87604513
All text-side matrices of the dual-stream blocks are kept in BF16, following
the Lens conversion profile. The 24 MLP down projections (mlp*.w2) are kept
in BF16 because their input width (6826) is not divisible by the ConvRot group
size. The modulation, text adapter, embedders, pixel head, final layer,
normalization and bias tensors are also kept in BF16.
This is a test conversion. Check output quality for your own prompts and workflows before relying on it.
License
The original model is released under the Apache License 2.0. A copy of the license is included in this repository.