Iris-3B Turbo landed on Hugging Face on October 12, and it is one of the more interesting free image models of the autumn. Sperid Labs distilled its 3-billion-parameter Iris-3B down to a 4-step model that makes a 1024×1024 image in about 0.7 seconds on an H100 (their number). The whole thing is Apache 2.0, so you can use the outputs and the weights commercially.
What makes Iris different is that it skips the VAE entirely. Most image generators work in a compressed latent space and decode back to pixels at the end. Iris paints every pixel directly, and the same backbone, fine-tuned, also does depth maps and 4× image restoration. You can try all of it free in a browser today, no GPU needed.
Iris-3B Turbo at a Glance
- What it is: a 4-step distilled version of Iris-3B, a pixel-space text-to-image diffusion transformer (no VAE, no latents).
- What’s free: everything. Weights, training code and two hosted demos on Hugging Face ZeroGPU.
- License: Apache 2.0 for the model; the Qwen3-VL-4B text encoder it uses is also Apache 2.0.
- Bonus tools: the same repo ships a depth estimator and a 4× restoration/upscaler built on the same backbone.
- The catch: official code needs an NVIDIA CUDA GPU, the full repo is roughly 42 GB, and ComfyUI support is community-built and partly experimental.
What Is Iris-3B, and Why Does “Pixel-Space” Matter?
Iris-3B is a diffusion transformer with 8 dual-stream blocks, 16 single-stream blocks and a 4-block “pixel head” that turns each 16×16 patch straight back into RGB pixels. It was trained from scratch with rectified flow at 256, then 512, then 1024 px, followed by supervised fine-tuning, for 665K steps in total. Text understanding comes from a frozen Qwen3-VL-4B-Instruct encoder, which downloads automatically on first run.
For creators, the practical promise of pixel-space generation is fine texture. Latent decoders have a known habit of smearing skin pores, film grain and tiny text into a plasticky average. Sperid Labs’ pitch is that “nothing is lost to a lossy, texture-biased latent.” Their arXiv paper and project page go deep on the method. Our read: it won’t dethrone the biggest models, but it is a clean, permissive base to build on.
What Turbo Changes: 100 Steps Down to 4
The base Iris-3B defaults to 100 denoising steps with classifier-free guidance, which works out to about 200 forward passes per image. Turbo bakes the guidance into the weights and samples in exactly 4 steps, one forward pass each. It is a full fine-tune of all 3B parameters (not a LoRA), distilled with Trajectory Distribution Matching, and ships as a ~6 GB bf16 file in the turbo/ folder.
Because the schedule is part of the model, --steps, --cfg-scale and --negative-prompt do nothing on Turbo. You control the prompt, seed and size. It was distilled in seven aspect buckets around one megapixel: 1024×1024, 752×1344, 896×1344, 896×1200 and their transposes. Stick to those.
On Sperid Labs’ own benchmark runs, Turbo scores 0.820 on GenEval and 86.68 on DPG-Bench versus 0.798 and 86.52 for the base model, so prompt following holds up. It drops on long rendered text (LongText-Bench 0.804 vs 0.857) and on diversity. These are vendor-measured numbers using official scorers; we have not seen independent leaderboard results yet, and a 0.02 GenEval gap is close to noise.
One Backbone, Three Jobs: Depth Maps and 4× Restoration
This is the part most coverage will skip. The same pixel-space prior, fine-tuned for 10K steps per task with no architecture changes, becomes a monocular depth estimator and an image restoration/upscaling model. Both run in a single forward pass with an empty prompt, so the text encoder is not even loaded.
- Depth: relative (affine-invariant, not metric) depth, with a reported AbsRel of 0.071 averaged across five zero-shot benchmarks using the Marigold V2 protocol. If you use depth maps for parallax, relighting or ControlNet, compare it against Marigold V2, which we covered last month.
- Restoration and upscaling: one-step 4× restoration trained with the HYPIR recipe. Large outputs are tiled into overlapping 1024×1024 windows.
Iris-3B Turbo vs Other Free Fast Image Models
| Model | Size | Steps | License | Free demo? | ComfyUI |
|---|---|---|---|---|---|
| Iris-3B Turbo | 3B + 4B text encoder | 4 | Apache 2.0 | Yes (HF Space) | Community nodes, partly experimental |
| Qwen-Image-2.1-Turbo | 7B | 8 | Read the fine print (see our review) | Yes | Yes |
| FLUX.2 [klein] 4B | 4B | 4 | Apache 2.0 | Yes | Native |
| Z-Image-Turbo | 6B | 8 | Apache 2.0 | Yes | Native |
If you need native ComfyUI today and the widest ecosystem, FLUX.2 [klein] 4B or Z-Image-Turbo are still the easier installs (we compared Klein in our Nano Banana 2.1 piece). Yesterday’s Qwen-Image-2.1-Turbo review covers editing, which Iris does not do.
Try Iris-3B Turbo Free Today
No GPU (fastest path): open the official Iris-3B Turbo demo Space, type a prompt and hit generate. For depth maps and upscaling, use the main Iris-3B Space, which has text-to-image, depth and upscaling tabs. Both run on free ZeroGPU, so you get a daily quota and may queue at peak times; a free Hugging Face login raises the quota.
Run it locally (NVIDIA GPU, Python 3.11+, PyTorch 2.7.1+):
git clone https://github.com/speridlabs/iris-3b.git
cd iris-3b
pip install -e .
hf download speridlabs/iris-3b --include "turbo/*" --local-dir iris-3b # ~6 GB
python scripts/sample.py --checkpoint iris-3b/turbo \
--prompt "a red fox sleeping in fresh snow, golden hour"- Downloads: Turbo ~6 GB, base model ~12 GB, depth ~12 GB, upscaler ~12 GB, plus the Qwen3-VL-4B text encoder on first run. Grab only the folder you need.
- Hardware: Sperid Labs does not publish a minimum VRAM figure. Their own numbers: all three models loaded together take ~45 GB; with
IRIS_OFFLOAD=1the local Gradio app (pip install -e ".[ui]"thenpython demo/app.py) peaks around 20 GB. Our estimate for Turbo alone (6 GB weights plus the 4B encoder) is a 16–24 GB card; treat that as unverified. - Smaller cards: Rebel AI’s ComfyUI-Iris-3B nodes load a 3.06 GB W4A8 checkpoint, and their GGUF builds run from 1.29 GB (Q3_K_S) to 2.46 GB (Q6_K) with city96’s ComfyUI-GGUF installed alongside. The authors flag GGUF as experimental and say smaller files do not guarantee lower VRAM.
- LoRA route: the community iris-3b-turbo-lora is a 460 MB 4-step adapter for the base model, with its own demo Space. It scores below the official Turbo on every benchmark Sperid Labs ran.
Everything else, including the batch-prompt flag and the depth and upscale scripts, lives on the Iris-3B model card.
Honest Trade-offs
- Text rendering: short headlines look good, but dense or long text is less reliable on Turbo than on the base model, by the developers’ own admission.
- Less variety: different seeds for the same prompt give more similar images. If you explore by rerolling, use the base model.
- Flat-background speckle: faint grain can show up on vector-style flat backgrounds, which matters for logo and sticker work.
- No Mac or AMD path in the official code: it requires CUDA. ComfyUI has no native Iris support yet; you depend on community nodes.
- Resolution: about one megapixel. For print, run the result through the built-in upscaler.
- Unverified speed claims: 0.7 s per image is on an H100 at batch 4. Expect several seconds on a consumer card.
Workflow Tips for Iris-3B Turbo
- Draft on Turbo, finish on base. Lock composition with fast 4-step drafts, then rerun your favourite seed on the 100-step base model when you need long text or more detail.
- Write plain descriptive sentences. The model card recommends plain English descriptions over tag soup, and the Qwen3-VL encoder rewards it.
- Use the trained aspect buckets. 752×1344 for vertical social, 1344×752 for YouTube-style thumbnails.
- Chain the three models. Generate, upscale 4× with
scripts/upscale.py, then pull a depth map withscripts/depth.pyfor parallax or relighting, all from one Apache 2.0 repo. - Batch from a text file. Use
--txt-file prompts.txtwith one prompt per line to run a whole moodboard overnight.
Watch our Minimax H3 vs. Flux 3 real estate interiors test, and subscribe to the Creative AI Show YouTube channel for weekly head-to-heads. More free tools live on Creative AI Show.
Iris-3B Turbo FAQ
Is Iris-3B Turbo free?
Yes. The weights, code and training scripts are free on Hugging Face and GitHub, and you can generate images at no cost in the official Hugging Face Space, subject to the daily ZeroGPU quota.
Can I use Iris-3B commercially?
Yes. Iris-3B and Iris-3B Turbo are released under Apache 2.0, and the Qwen3-VL-4B-Instruct text encoder is also Apache 2.0. Still review outputs for likenesses and trademarks.
What are Iris-3B Turbo’s system requirements?
The official code needs an NVIDIA CUDA GPU, Python 3.11+ and PyTorch 2.7.1+. Turbo is a ~6 GB download plus the 4B text encoder; no official minimum VRAM is published, and the developers’ offloaded demo peaks around 20 GB. Community W4A8 and GGUF builds are smaller but experimental.
Does Iris-3B work in ComfyUI?
Not natively yet. Rebel AI’s ComfyUI-Iris-3B custom nodes support the FP32 and W4A8 checkpoints, with an experimental GGUF add-on.
What is the difference between Iris-3B and Iris-3B Turbo?
Iris-3B is the 100-step base model with adjustable guidance and negative prompts. Turbo is a 4-step distilled version that is far faster and slightly better at prompt following on vendor benchmarks, but weaker on long text and seed variety.
Accurate as of October 13, 2026. Model versions, demo quotas and community ComfyUI support move fast.
No responses yet