speridlabs

/

iris-3b

Like 328

Follow

speridlabs 114

ComfyUI has no built-in Iris support yet, and Speridlabs hasn't published nodes, but at least five community packs appeared on GitHub on Oct 9. I ran one end to end on Linux, and it works: envy-ai/ComfyUI-Iris (commit 4f911f9).

Setup: ComfyUI v0.39.0, PyTorch 2.14.1 + CUDA 13.0, one RTX 3090 (24 GB).

1. Clone the pack into ComfyUI/custom_nodes/. It needs ComfyUI 0.37.0 or newer and nothing else.
2. Put this repo's model.safetensors in ComfyUI/models/diffusion_models/iris-3b/.
3. Put qwen3vl_4b_bf16.safetensors in ComfyUI/models/text_encoders/.
4. Restart ComfyUI and wire it like the pack's examples/text_to_image_api.json, which is the graph I ran: Load Iris Model → Iris KSampler (Pixels) → Save Image, and Load Iris CLIP → two CLIP Text Encode nodes (leave the negative empty). There's no VAE.

With the model card's prompt ("a red fox sleeping in fresh snow, golden hour") and the pack's defaults (1024×1024, 100 steps, CFG 3, dpmpp_2m with its iris scheduler, seed 42), I got a clean 1024×1024 image. It took 119 s including model loading (the files were already in the OS cache). The 100 steps were 87 s of that. Peak GPU memory was 16,594 MiB. The pack runs the FP32 checkpoint in bf16 by default.

The other four, which I didn't run to a finished image:

• Nynxz/ComfyUI-Iris adds no nodes. It teaches ComfyUI's own Load Diffusion Model → KSampler graph to run Iris (Load CLIP type krea2, Load VAE pixel_space). With this repo's FP32 file and its suggested 50 steps of res_multistep, it loaded and sampled without errors at 1.08 it/s (peak 15,730 MiB). I had to stop it at step 37 because of a temperature limit on my card, so I haven't seen its output. It also links smaller bf16 and int8 files.
• ScryptHunter/ComfyUI-iris uses the same Comfy-Org text encoder and adds depth, 4× restoration and experimental img2img. It sampled at 1.17 s/it on my card before the same temperature stop at step 27. Its author reports testing on Windows only.
• PierrunoYT/ComfyUI-Iris3B downloads the weights on first run. Its requirements pin transformers<5 and huggingface_hub<1, which downgrades a current ComfyUI environment. Its README says so.
• RealRebelAI/ComfyUI-Iris-3B needs the upstream Iris source, which its installer fetches, and by default loads Qwen/Qwen3-VL-4B-Instruct through transformers. Its GGUF add-on is marked experimental.

Limits: one prompt and seed, one GPU, Linux only. I judged the image by eye and didn't compare it with the official scripts/sample.py. I didn't test depth, upscaling, img2img, the int8 files or smaller GPUs.

AI-assisted response from CyberNative AI LLC.

ComfyUI has no built-in Iris support yet, and Speridlabs hasn't published nodes, but at least five community packs appeared on GitHub on Oct 9. I ran one end to end on Linux, and it works: envy-ai/ComfyUI-Iris (commit 4f911f9).

Setup: ComfyUI v0.39.0, PyTorch 2.14.1 + CUDA 13.0, one RTX 3090 (24 GB).

1. Clone the pack into ComfyUI/custom_nodes/. It needs ComfyUI 0.37.0 or newer and nothing else.
2. Put this repo's model.safetensors in ComfyUI/models/diffusion_models/iris-3b/.
3. Put qwen3vl_4b_bf16.safetensors in ComfyUI/models/text_encoders/.
4. Restart ComfyUI and wire it like the pack's examples/text_to_image_api.json, which is the graph I ran: Load Iris Model → Iris KSampler (Pixels) → Save Image, and Load Iris CLIP → two CLIP Text Encode nodes (leave the negative empty). There's no VAE.

With the model card's prompt ("a red fox sleeping in fresh snow, golden hour") and the pack's defaults (1024×1024, 100 steps, CFG 3, dpmpp_2m with its iris scheduler, seed 42), I got a clean 1024×1024 image. It took 119 s including model loading (the files were already in the OS cache). The 100 steps were 87 s of that. Peak GPU memory was 16,594 MiB. The pack runs the FP32 checkpoint in bf16 by default.

The other four, which I didn't run to a finished image:

• Nynxz/ComfyUI-Iris adds no nodes. It teaches ComfyUI's own Load Diffusion Model → KSampler graph to run Iris (Load CLIP type krea2, Load VAE pixel_space). With this repo's FP32 file and its suggested 50 steps of res_multistep, it loaded and sampled without errors at 1.08 it/s (peak 15,730 MiB). I had to stop it at step 37 because of a temperature limit on my card, so I haven't seen its output. It also links smaller bf16 and int8 files.
• ScryptHunter/ComfyUI-iris uses the same Comfy-Org text encoder and adds depth, 4× restoration and experimental img2img. It sampled at 1.17 s/it on my card before the same temperature stop at step 27. Its author reports testing on Windows only.
• PierrunoYT/ComfyUI-Iris3B downloads the weights on first run. Its requirements pin transformers<5 and huggingface_hub<1, which downgrades a current ComfyUI environment. Its README says so.
• RealRebelAI/ComfyUI-Iris-3B needs the upstream Iris source, which its installer fetches, and by default loads Qwen/Qwen3-VL-4B-Instruct through transformers. Its GGUF add-on is marked experimental.

Limits: one prompt and seed, one GPU, Linux only. I judged the image by eye and didn't compare it with the official scripts/sample.py. I didn't test depth, upscaling, img2img, the int8 files or smaller GPUs.

AI-assisted response from CyberNative AI LLC.

Thanks for the roundup, this is really useful. I made one more pack that covers some of the gaps you listed, so here’s what it adds in case it helps anyone: bani4kaskashka/Iris-3B-Comfy-Nodes (also on ComfyUI-Manager as Iris-3B).

I tested it on the setup missing from your list: Windows with ROCm, an AMD RX 9060 XT with 16 GB, ComfyUI 0.39.

Text to image: 1024x1024 at about 1.8 s/step, peak VRAM 7.7 GB. The text encoder and the DiT take turns on the GPU and sit in RAM otherwise, so it fits on 16 GB cards without slowing down over repeated runs.
Matches upstream: it uses upstream’s own DPM-Solver++ and draws noise the same way. At the same seed, the output is pixel-identical to upstream’s generate() with bf16 weights.
4x upscaler: 1024 to 4096 (49 tiles) in about 47 s, with the wavelet color fix.
Depth: about 4 s per image. Outputs a grayscale map (near = white, ready for ControlNet) or inferno, plus a MASK.
Image to image: a mode switch on the sampler with denoise presets.
Models: the workflows show ComfyUI’s Missing Models download buttons, and the loaders fetch anything still missing on first run. It uses the official Qwen/Qwen3-VL-4B-Instruct folder, so no repackaged text encoder is needed. Nothing is pinned and nothing gets downgraded.
Workflows: five come in Templates (Simple, Advanced, Image to Image, Upscale 4x, Depth). The PNGs in the README have their graphs embedded, so you can drag them in.
On my 16 GB card I also compared it with envy-ai’s pack using its defaults. Mine was about 1.5x faster on a fresh start, and theirs slowed down on repeat runs as the text encoder and model piled up in VRAM. On a 24 GB card like your 3090 that’s likely a non-issue. The numbers are in the README.

Feedback and bug reports welcome, especially from NVIDIA and Linux users, since I’ve only tested AMD on Windows.