README.md

metadata

license: apache-2.0
base_model:
  - speridlabs/iris-3b
  - Qwen/Qwen3-VL-4B-Instruct
pipeline_tag: text-to-image
library_name: fusor-ai
tags:
  - webassembly
  - wasm
  - on-device
  - text-to-image
  - depth-estimation
  - image-to-image
  - super-resolution

Iris-3B, compiled for the browser

speridlabs' Iris-3B, a pixel-space text-to-image model with a four-step Turbo, a depth estimator and a four-times restorer, compiled with the Qwen3-VL-4B text encoder it reads its prompts through into five fusor-ai bundles: the weights the browser runtime of iris-wasm loads to turn a prompt or a photo into an image on the device's CPU. Try it in the Space fusorlabs/iris-wasm.

• Sources: speridlabs/iris-3b at revision adc05b909fb05d4276d26aae5a52b36406b2ef26 and Qwen/Qwen3-VL-4B-Instruct at revision ebb281ec70b05090aa6165b016eac8ec08e71b17, compiled with the mix profile into five bundles, because one WebAssembly memory holds at most 4 GiB and the page runs the text encoder and one transformer at a time:
  • text/: Qwen3-VL's language model up to the last layer Iris reads, and its tokenizer; the feed-forward matrices 4-bit, every other large matrix 7-bit (3105 MB).
  • base/, turbo/, depth/, upscaler/: one transformer each (the base model, its four-step distillation, the depth estimator and the restorer); the feed-forward matrices of the blocks 4-bit, the other large matrices 7-bit, the layers that read straight into or out of pixels in f32 (2558 MB each). The depth and the restorer bundles hold the empty prompt's states.
• Format: each manifest.json lists every tensor and part; each part is named by its BLAKE3 hash, which the runtime checks as it loads.
• Reproduce it: just bundles in iris-wasm downloads the checkpoints and compiles the same bundles.

This is not a PyTorch or safetensors checkpoint: it only loads in the iris-wasm runtime.

The models are Apache-2.0 (LICENSE, NOTICE).