Iris3b

Iris-3B in one workflow: make pictures, read depth, or restore and enlarge. Iris paints directly in pixel space (no VAE), at full-quality weights. PICK A MODE with slot1: 0 (default) - Text to image: describe the picture in plain language (subject, materials, light, framing; about 295 tokens are read). No negative prompt on purpose - Iris is trained to work with an empty one. Uses the official 100-step reference sampler. Controls: steps, guidance (3 = reference), seed, width/height (about 1 megapixel, multiples of 16, e.g. 1024x1024, 1216x832), slot2 = sampling shift (4 = reference). 1 -…Iris-3B in one workflow: make pictures, read depth, or restore and enlarge. Iris paints directly in pixel space (no VAE), at full-quality weights. PICK A MODE with slot1: 0 (default) - Text to image: describe the picture in plain language (subject, materials, light, framing; about 295 tokens are read). No negative prompt on purpose - Iris is trained to work with an empty one. Uses the official 100-step reference sampler. Controls: steps, guidance (3 = reference), seed, width/height (about 1 megapixel, multiples of 16, e.g. 1024x1024, 1216x832), slot2 = sampling shift (4 = reference). 1 - Depth map: attach an image (image1) and set slot1 to 1. You get a grey-scale relative depth picture, dark = near, bright = far. slot3 = detail (megapixels the image is resized to, 0.6 default, up to 1.0). 2 - Upscale x4 / restore: attach an image (image1) and set slot1 to 2. Cleans up and enlarges 4x. Works best on small or soft images; the image is first resized to about 0.1 megapixel (slot4 sets this, up to 0.25), so the result is roughly 1300 pixels wide. Made of: Iris-3B by Speridlabs (Apache-2.0) with its depth model and upscaler LoRA, bf16, plus a Qwen3-VL 4B text encoder, running through the ComfyUI-Iris loader. The first run on a cold machine downloads about 21 GB and can take several minutes.もっと見る