Apache-2.0image-genRuns locally3B
Iris-3B: Pixel-Space Text-to-Image (and Turbo)
3B pixel-space text-to-image model (no VAE), with a 4-step Turbo distill in the same repo.
- License
- Apache-2.0
- Parameters
- 3B
- HF likes
- 200 (2026-10-13)
- First seen
- 2026-10-05
TL;DR. Iris-3B is a image-gen model from speridlabs. Parameters: 3B.License: Apache-2.0. It is listed under Midjourney.
| Developer | speridlabs |
|---|---|
| Type | image-gen |
| Architecture | Pixel-space diffusion transformer (no VAE); frozen Qwen3-VL-4B-Instruct text encoder |
| Parameters | 3B |
| Modalities | image, text |
| License | Apache-2.0 — Apache-2.0 on the card, GitHub LICENSE, and NOTICE as of 2026-10-13. The frozen Qwen3-VL-4B-Instruct encoder is separately Apache-2.0. Not legal advice. |
| Hardware | Not yet verified — Card requires an NVIDIA GPU with CUDA. T2I weights about 12 GB; turbo/ about 6 GB. Demo with all three checkpoints cites ~45 GB, or ~20 GB peak with IRIS_OFFLOAD=1. Not independently measured. (not independently tested) |
| Hugging Face likes | 200 (as of 2026-10-13) |
| Model card | https://huggingface.co/speridlabs/iris-3b |
| First seen | 2026-10-05 |
Iris-3B is a 3-billion-parameter pixel-space diffusion transformer from speridlabs. Official hosts: Hugging Face speridlabs/iris-3b, GitHub speridlabs/iris-3b, and the project page speridlabs.com/research/iris.
This identification page hangs off open Midjourney-class alternatives. That mapping is editorial. We have no sourced bake-off against Midjourney.
Iris-3B Turbo (4-step)
The same repository’s turbo/ folder is Iris-3B Turbo: a 4-step distilled version. The Hugging Face card was last modified 2026-10-12. The card says a 1024×1024 image takes about 0.7 s on one H100 (batch 4). That timing is upstream, not a forkalt measurement.
A community LoRA (iris-3b-turbo-lora) is thanked on the card; it is not the official Turbo checkpoint.
What we verified (2026-10-13)
| Field | Value | Source |
|---|---|---|
| HF likes | 200 | HF API |
| Card created | 2026-10-05 | HF createdAt |
| GitHub created | 2026-10-05 | GitHub API |
| GitHub stars / forks | 131 / 12 | GitHub API |
| License | Apache-2.0 | HF card, GitHub license, NOTICE |
| Parameters | 3B + frozen 4B text encoder | Model card |
| Default T2I sampling | CFG 3, 100 steps, ~1024×1024 | Card tips table |
| Turbo | 4-step distill in turbo/ |
Card; repo last push 2026-10-12 |
Benchmarks in the table on this page are self-reported from the official card. forkalt has not reproduced them. Compare only with Iris-3B vs Qwen-Image-2.1-Turbo for license and architecture facts, not quality ranking.
What we did not invent
- Peak VRAM for a single T2I run is not yet verified.
- First-party ComfyUI support is not yet verified.
- We do not claim visual quality relative to Midjourney or Qwen-Image.
Benchmarks
Quoted from the cited source. forkalt has not independently reproduced these numbers.
| Name | Score | Source |
|---|---|---|
| GenEval (Iris-3B) | 0.798 (self-reported) | https://huggingface.co/speridlabs/iris-3b |
| GenEval (Iris-3B Turbo) | 0.820 (self-reported) | https://huggingface.co/speridlabs/iris-3b |
| DPG-Bench (Iris-3B) | 86.52 (self-reported) | https://huggingface.co/speridlabs/iris-3b |
| DPG-Bench (Iris-3B Turbo) | 86.68 (self-reported) | https://huggingface.co/speridlabs/iris-3b |
Deploy options
Comparisons
FAQ
What is “pixel-space” here?
The official card says Iris-3B generates raw RGB pixels with no VAE and no latent space. Default sampling in the tips table is CFG 3 and 100 steps at about 1024×1024.
What is Iris-3B Turbo?
A 4-step distilled checkpoint in the `turbo/` folder of the same Hugging Face repo. The card (last modified 2026-10-12) calls it a full fine-tune of all 3B parameters, not a LoRA, with classifier-free guidance baked in. Official sample code uses `scripts/sample.py --checkpoint iris-3b/turbo`.
Does it have first-party ComfyUI support?
Not yet verified from official docs. The GitHub README documents Python sampling and a Gradio demo. A GitHub issue titled “Comfyui Port” was opened 2026-10-10, which reads as a request rather than a shipped integration. Community Spaces exist; we do not treat those as first-party.
What is the license?
Apache-2.0 on the Hugging Face card and GitHub repository as of 2026-10-13. Not legal advice.
Sources & verification
- https://huggingface.co/speridlabs/iris-3b
- https://github.com/speridlabs/iris-3b
- https://speridlabs.com/research/iris
- https://huggingface.co/speridlabs/iris-3b/blob/main/assets/turbo.webp
License notes are not legal advice. Read the original LICENSE. Trademark & disclaimer.