I made ComfyUI nodes for Iris-3B that supports - text to image, 4x upscaler, depth and img2img, works on a 16 GB AMD card
retained text
Iris-3B from Sperid Labs came out this week. It's a 3B model that works directly in pixels with no VAE, and it also ships a 4x upscaler and a depth model. ComfyUI doesn't support it natively yet, so I made a node pack that covers all models and use cases. **Repo:** [https://github.com/bani4kaskashka/Iris-3B-Comfy-Nodes](https://github.com/bani4kaskashka/Iris-3B-Comfy-Nodes) Also in ComfyUI-Manager: search **Iris-3B**. Free, Apache-2.0. **What's in it** * **Simple sampler:** dropdowns for resolution (1 MP sizes from 21:9 to 9:21), quality (30, 50 or 100 steps), prompt adherence, seed and batch * **Advanced sampler:** every setting (steps, cfg, shift, solver order, CFG interval, any size) * **Image to image:** a mode switch on both samplers, with subtle, medium and strong presets * **4x upscaler:** works on any image, 1024 to 4096 in about 47 s * **Depth:** about 4 s per image, grayscale (ready for ControlNet) or inferno, plus a mask * **Five workflows** in Templates. The images in the README have their workflows embedded, so you can drag them straight into ComfyUI. **Setup** * Open a workflow and ComfyUI shows download buttons for any missing models. * The loaders also download anything still missing on first run. * It uses the official Qwen3-VL-4B text encoder folder, so you don't need a repackaged file. * No extra pip installs, and nothing in your ComfyUI gets downgraded. **Speed and VRAM** (RX 9060 XT, 16 GB, Windows, ROCm) * About 1.8 s/step at 1024x1024 * Peak VRAM 7.7 GB. The text encoder and the model take turns on the GPU, so it doesn't slow down over repeated runs on 16 GB cards. * The README has a comparison table if you're curious. The fox girl's sign text is straight from the model, no inpainting. I generated six seeds and every one spelled "IRIS 3B FOR COMFY BY BUCKY" correctly, which surprised me for a 3B model. I've only tested on AMD and Windows so far, so reports from NVIDIA and Linux users would really help. Bugs and feature requests go to GitHub issues. Images included should be drag and drop to comfy for the workflow.
document fetched 2026-10-11 00:31 · canonical https://reddit.com/r/StableDiffusion/comments/1x2ksba/i_made_comfyui_nodes_for_iris3b_that_supports