# Iris-3B (slim)

> Iris-3B is an Apache-2.0 open-weight text-to-image model from Speridlabs that generates pixels directly, with no VAE. The release adds a 4-step Turbo version and fine-tunes for monocular depth and 4x image restoration.

- Full: https://www.anchorterminal.com/tools/iris-3b.md (~6,000 tokens) · this version ~1,680 tokens · JSON https://www.anchorterminal.com/tools/iris-3b.json · canonical https://www.anchorterminal.com/tools/iris-3b
- Index: https://www.anchorterminal.com/llms.txt · API: https://www.anchorterminal.com/api/v1/index.json · Updated: 2026-10-15
- Licence: CC BY 4.0, credit Anchor Terminal (anchorterminal.com) and link this page when you pass anything on

**D · 50.9/100 · rank #826 of 1094 · #8 in Image generation · not agent-ready · confidence medium**


Assessment: Code and weights are Apache-2.0, and one Python package runs text-to-image, a 4-step Turbo, depth and 4x restoration from a documented CLI. It is ten days old, with no PyPI package, version tags, changelog or CI, needs an NVIDIA GPU and about 12 GB of weights, and has no server or API of its own.

## Facts

- Kind: Model API · vendor: Speridlabs · category: Image generation · legal entity: Speridlabs · provenance 44/100
- Auth: None · pricing: Free · x402: no · licence: Apache-2.0 (code and weights)
- Probe metrics: not measured yet (probes haven't run)
- Surface graded: The `iris-3b` Python package and its CLI scripts (`sample.py`, `depth.py`, `upscale.py`) run on the owner's GPU, read from the repository, the model card and the paper on 15 October 2026. Nothing was installed or run
- Checkpoints: Base text-to-image (fp32, about 12 GB), `turbo/` 4-step distilled (bf16, about 6 GB), `depth/` relative monocular depth and `upscaler/` 4x restoration (about 12 GB each), all in speridlabs/iris-3b on Hugging Face
- Model: 3B-parameter diffusion transformer in pixel space, patch size 16, with a frozen Qwen3-VL-4B-Instruct text encoder. Rectified flow, trained from scratch for 665K steps
- Output: About one megapixel. CLI default 1024x1024. Turbo was distilled at seven aspect buckets (1024x1024, 752x1344, 896x1344, 896x1200 and their transposes)
- Hardware: NVIDIA GPU with CUDA per the model card. The demo README gives about 45 GB of GPU memory for all three models, or about 20 GB peak with `IRIS_OFFLOAD=1`. Memory for text-to-image alone was not stated
- Interfaces: CLI scripts and Python (`iris3b.sampling.generate`, `generate_few_step`, `iris3b.downstream.depth.DepthPredictor`, `iris3b.downstream.restoration.Restorer`). No HTTP server, OpenAPI or MCP. Not in diffusers or ComfyUI officially; community ComfyUI packs and GGUF and INT8 conversions exist
- Hosted demos: Gradio Spaces speridlabs/iris-3b (generate, depth, upscale) and speridlabs/iris-3b-turbo on Hugging Face ZeroGPU. Their Gradio API lists typed endpoints (`/run`, `/run_depth`, `/run_upscale`). Speridlabs doesn't document them as an API
- Image editing: Not supported. The project lead wrote on 11 October 2026 that editing is not on the roadmap for this base model (GitHub issue 1)
- Benchmarks (vendor): Iris-3B GenEval 0.798, DPG-Bench 86.52, LongText-Bench EN 0.857, OneIG-Bench EN 0.540. Turbo 0.820, 86.68, 0.804 and 0.516. Vendor-reported at 1024x1024
- Releases: Base weights 5 October, depth and restoration 8 October, Turbo 12 October 2026. Paper on arXiv 7 October. Package version 0.1.0, no tags
- Adoption: On 15 October 2026 the model repository showed 267 likes and 1,132 downloads in the last month, with 12 derived models, and the GitHub repository 175 stars
- Age: Ten days old on 15 October 2026. The first commit and the first weights are dated 5 October 2026

- Scores: Reliability 45, Performance pending, Schema & documentation 47, Agent ergonomics 54, Security & auth 44, Payments & pricing 60, Task success pending, Maintenance & community 70, Transparency & trust 47 · total over the 7 assessed categories
- Why: Reliability, Read as local software, since Iris-3B is a model the owner runs. · Schema & documentation, Read for a model the owner runs through a CLI and a Python package. · Agent ergonomics, Read as a local image model an agent calls through the CLI or Python. · Security & auth, Read as software the owner runs. · Payments & pricing, Free Apache-2.0 software with nothing to buy, so 20 + 20 + 20 for pricing, free use and no sign-up. · Maintenance & community, Read for an open-weight model ten days old. · Transparency & trust, Apache-2.0 for code and weights, with a NOTICE file, a third-party licence table and the full training and distillation recipe.
- Sources: 12, open questions: 5, both in the full twin
- Capabilities: image.generate, image.upscale
- JSON: https://www.anchorterminal.com/api/v1/tools/iris-3b.json
- Verify (for the vendor): the badge `https://www.anchorterminal.com/badges/iris-3b.svg` or a link to https://www.anchorterminal.com/tools/iris-3b from a page on speridlabs.com or one of its subdomains, or the README of github.com/speridlabs/iris-3b, then `POST https://www.anchorterminal.com/api/v1/verify` `{"slug", "url"}` or `verify_listing` at /mcp; re-checked weekly, no effect on the grade. Snippets in the full twin.

## Before you call it

1. Check for an NVIDIA GPU with CUDA and Python 3.11 or newer before downloading. The base weights are about 12 GB, and Qwen3-VL-4B downloads on the first run
2. Exclude `depth/*`, `upscaler/*` and `turbo/*` in `hf download` unless those models are needed. Each of the first two adds about 12 GB
3. Don't pass `--steps`, `--cfg-scale`, `--negative-prompt`, `--order` or `--shift` with a Turbo checkpoint. The script exits with an error
4. Pass `--seed` to reproduce an image, and set `--outdir`, since files are named after the prompt and a repeat run overwrites them
5. Load only exported directories (`model.safetensors` and `config.yaml`). A training `.pth` is read with `torch.load(weights_only=False)`, so never load one from an untrusted source

## Connect

```bash
git clone https://github.com/speridlabs/iris-3b.git && cd iris-3b && pip install -e .
hf download speridlabs/iris-3b --local-dir iris-3b --exclude "depth/*" "upscaler/*" "turbo/*"
python scripts/sample.py --checkpoint iris-3b --prompt "a red fox sleeping in fresh snow, golden hour"
```

## Similar tools

| Tool | Grade | Score | Shared capabilities | Slim |
| --- | --- | --- | --- | --- |
| fal image models | B | 65.2 | image.generate, image.upscale | https://www.anchorterminal.com/tools/fal-image.min.md |
| Bria API | B | 63.7 | image.generate, image.upscale | https://www.anchorterminal.com/tools/bria.min.md |
| Photoroom API | C | 56 | image.generate, image.upscale | https://www.anchorterminal.com/tools/photoroom.min.md |
| Melius | C | 54.1 | image.generate, image.upscale | https://www.anchorterminal.com/tools/melius.min.md |
| Recraft API | D | 52.8 | image.generate, image.upscale | https://www.anchorterminal.com/tools/recraft.min.md |


## Panel reviews (0, desk reviews from public material, no calls made)