Iris-3B GGUFs — Rebel AI
GGUF quantizations of SperiLabs Iris-3B, a 3B-parameter pixel-space image generator that produces RGB images directly without a VAE.
By Rebel AI · ComfyUI nodes on GitHub
Intended GGUF variants
| Quantization |
Relative precision |
Notes |
| Q3_K_S |
Smallest |
Most aggressive memory savings; validate image quality. |
| Q4_K_S |
Balanced |
Starting point for experiments. |
| Q5_K_S |
Higher |
More weight precision. |
| Q6_K |
Highest of these |
Largest memory demand. |
Availability and exact filenames should be checked in this repository's Files tab. The GGUF model contains Iris diffusion weights, not the Qwen3-VL encoder. A separate text encoder is required.
Run in ComfyUI
- Install ComfyUI-Iris-3B and its upstream Iris source dependencies.
- Install city96/ComfyUI-GGUF.
- Install the optional, isolated
gguf_addon/ from the ComfyUI repository as a separate sibling custom-node folder. Do not overwrite a working W4A8/FP32 installation.
- Copy the GGUF you want to run into
ComfyUI/models/checkpoints/Iris-3B/.
- In ComfyUI, connect
Iris 3B GGUF Checkpoint to the existing Iris 3B Qwen3-VL Encode node, then connect both checkpoint and conditioning to Iris 3B GGUF Sampler (Experimental) → Save Image.
Recommended reference settings: 1024×1024 (or native ~1 megapixel), 100 steps, CFG 3.0, shift 4.0, order 2. The original model card showcases outputs near 1MP; early 512px trials exhibited 16px grid artifacts, while a 1024px W4A8 test produced clean output.
GGUF support is experimental. The ComfyUI addon has code-level validation, but real checkpoint execution, low-VRAM behavior and output quality are not yet independently validated for each GGUF format. Please report reproducible issues on the GitHub repository.
Architecture and compatibility
Iris-3B is a pixel-space rectified-flow diffusion model, not an LLM. Standard llama.cpp chat applications will not run these GGUFs. The adapter constructs the official Iris diffusion transformer and uses city96's GGUF tensor/dequantization backend for the model weights.
This base checkpoint is text-to-image, not a general prompt-based image editor. SperiLabs publishes separate fine-tuned weights for image restoration/upscaling and depth estimation in their official release; use the task-specific model for those features.
Credits and license
Original model, training and code: SperiLabs / iris-3b · License: Apache 2.0. GGUF conversion and ComfyUI integration: Rebel AI. GGUF loader infrastructure: city96/ComfyUI-GGUF.
These quantizations are derivatives of the original research model. Review the original model card for intended uses, limitations, and citation information.
