Iris-Z-small, quantized

INT8 ConvRot and W4A8 builds of Iris-Z-small by SperidLabs - one model that paints images, reads their depth, and repaints a scene from its depth, built on FLUX.2 [klein] 9B.

Made for Forge Neo with the Iris-Z for Forge Neo extension.

File Format Size Picture vs bf16 Depth vs bf16
iris-z-small_int8_convrot.safetensors INT8 ConvRot 9.4 GB 26.7 dB 40.1 dB
iris-z-small_w4a8.safetensors W4A8 5.6 GB 22.7 dB 32.8 dB
original bf16 (speridlabs/iris-z) bf16 18.2 GB - -

PSNR against the bf16 checkpoint, same prompt and seed, 1024x1024, 50 steps.

The same prompt and seed with the bf16, INT8 ConvRot and W4A8 checkpoints

What was quantized

Only the 112 linear layers of the 8 double and 24 single transformer blocks, as in the published FLUX.2 [klein] 9B INT8 ConvRot builds. The input and output layers, the timestep embedder, the stream modulations and Iris-Z's modality_in stay in bf16. Both files hold the diffusion model only - the text encoder and VAE are separate.

Quantized with Merge Studio (FLUX.2 profile).

  • INT8 ConvRot is the closest to the original.
  • W4A8 is the smallest, and runs at full speed on RTX 30 series cards or newer - on older cards choose INT8.

How to use

  1. Install the Iris-Z for Forge Neo extension in Forge Neo
  2. Put one of these files in models/Stable-diffusion
  3. Add the Qwen3 8B text encoder (qwen_3_8b or qwen_3_8b_fp8mixed) to models/text_encoder and the FLUX.2 VAE (flux2-vae) to models/VAE
  4. Pick the iris_z UI preset - Euler, Flux2 schedule, 50 steps, CFG 4

Everything Iris-Z-small does works with both files:

Two pictures and their depth maps

A depth map repainted as an abandoned overgrown aqueduct, under the sea and in deep winter

Credits

  • Iris-Z by SperidLabs - the model (paper, code)
  • FLUX.2 by Black Forest Labs - the model Iris-Z is built on

License

The weights are derived from FLUX.2 and Iris-Z and stay under the FLUX Non-Commercial License, as the original Iris-Z weights.


Made with ❤️ for the Stable Diffusion community

Report Bug • Request Feature • ☕ Ko-fi

Downloads last month
616
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for eduardoabreu81/iris-z-small-quantized

Quantized
(1)
this model

Collection including eduardoabreu81/iris-z-small-quantized

Paper for eduardoabreu81/iris-z-small-quantized