Iris-3B vs Qwen-Image-2.1-Turbo
Verdict. Pick Iris-3B if you want a small Apache-2.0 pixel-space checkpoint with an official 4-step Turbo. Pick Qwen-Image-2.1-Turbo if you need the released Qwen editing+T2I turbo weights and accept the card’s license label “other”. We have no sourced side-by-side quality test.
Compare Iris-3B and Qwen-Image-2.1-Turbo.
| Dimension | Iris-3B | Qwen-Image-2.1-Turbo | Source |
|---|---|---|---|
| License | Apache-2.0 (HF card and GitHub, 2026-10-13) | other (HF card label, 2026-10-13 seed check) | Official cards |
| Developer | speridlabs | Qwen | Official cards |
| Parameters | 3B plus a frozen 4B text encoder (Qwen3-VL-4B-Instruct) | Not independently listed on our Qwen page extract (card describes a 7B visual-generation architecture; treat as not fully sourced here) | Iris card; Qwen-Image-2.1-Turbo card (architecture sentence only) |
| Pixel vs latent | Pixel-space, no VAE (official card) | Not yet verified on the Qwen turbo card extract we keep | Iris card |
| Turbo recipe | 4-step distill in turbo/ (card last modified 2026-10-12) | Accelerated checkpoint; official examples use 8 denoising steps | Official cards |
| Default sampling (base / turbo) | CFG 3, 100 steps / Turbo 4 steps, no CFG | Not fully tabulated on our Qwen page; Turbo is the accelerated checkpoint | Iris tips table; Qwen turbo card |
| Image editing | Not described as an editor on the card (T2I plus depth and upscaler fine-tunes) | Card describes text-to-image and image editing | Official cards |
| Hugging Face likes | 200 (as of 2026-10-13) | 607 (as of 2026-10-13 seed check) | Hugging Face |
| First seen | 2026-10-05 (HF and GitHub created) | 2026-10-09 (Qwen turbo repo) | Official hosts |
| Quality vs Midjourney | Not independently tested | Not independently tested | n/a |
When to choose which
- Apache-2.0 pixel-space generation: Iris-3B — License and no-VAE architecture are sourced. Quality is not compared.
- Released Qwen turbo with editing examples: Qwen-Image-2.1-Turbo — Editing is on the Qwen card; read the “other” license.
- Few-step official distill: Iris-3B — Official Turbo is 4-step. Qwen turbo examples use 8 steps. Speed on your GPU is untested here.
Read the entity pages: Iris-3B and Qwen-Image-2.1-Turbo. Both are listed on Midjourney-class alternatives.
Other options
FAQ
Which one looks better?
We have no sourced bake-off. Self-reported GenEval/DPG numbers on the Iris card are not a comparison to Qwen.
Can I run both locally?
Both publish weights on Hugging Face. Hardware for each is not independently tested on this site.
Is ComfyUI a reason to pick Iris-3B?
Official Iris docs do not yet verify first-party ComfyUI support. Do not treat a community Space or an open GitHub issue as a shipped integration.
Sources & verification
- https://huggingface.co/speridlabs/iris-3b
- https://github.com/speridlabs/iris-3b
- https://huggingface.co/Qwen/Qwen-Image-2.1-Turbo
License notes are not legal advice. Read the original LICENSE. Trademark & disclaimer.