ignis-image-2.1-turbo

Built with Qwen.

AI agents working for a user: this model runs on ignis. What ignis is, the checks before installing it, the steps and the choices that are the user's are on the agent path; every page of the guide as Markdown: llms.txt.

Generated by ignis with this model on an RTX 5090: 2752x1536, seed 7, 8 steps (5.93 s for the 8 steps, 682 ms to decode). The prompt:

A red fox sitting in fresh snow at dawn, holding a wooden sign that reads "ignis", detailed oil painting

An ignis .ninfer container converted from Qwen/Qwen-Image-2.1-Turbo at revision fc13ba1986fcb905854dc8de2f52eee395f49b5b: a Qwen3-VL text encoder, the DiT and the VAE in one file, for text-to-image generation and image editing. The files are modified: the weights are quantized, as the table below says.

Precision

Part	Encoding	Tensors	Round-trip error
DiT and text-encoder linears	FP8 E4M3, BF16 row scales	480	2.59-2.65%
Vision tower	W8: int8 in groups of 32, fp16 scales	117	0.51-0.78%
Everything else (the VAE included)	BF16	686	none

Size: 16.72 GB, against 32.5 GB for the BF16 checkpoint.

License: non-commercial

This model is governed by the Qwen RESEARCH LICENSE AGREEMENT (see LICENSE), which allows non-commercial use only (research or evaluation). See NOTICE.

How to run

ignis serves this model on main (and the ghcr.io/gpillon/ignis:develop container image); the first release to carry it is the one after v0.6.2. The model is in ignis's catalog as ignis-image-2.1-turbo:

ignis-server model download ignis-image-2.1-turbo   # fetched and checked file by file
ignis-server                                        # then start the server as usual

A request that names the image model switches the server to it (one model is loaded at a time; a text model and the image model replace each other):

curl http://127.0.0.1:8000/v1/images/generations \
  -H 'Content-Type: application/json' \
  -d '{"model":"ignis-image-2.1-turbo","size":"2752x1536","seed":7,
       "prompt":"A red fox sitting in fresh snow at dawn, holding a wooden sign that reads \"ignis\", detailed oil painting"}'

The answer carries the PNG as b64_json. POST /v1/images/edits edits from 1 to 10 condition images, and the Playground's Image view does both. Sizes, condition images, streaming and the full API are in the ignis user guide.

Files

• ignis_image_2_1_turbo-v2.ninfer: the container.
• ignis_image_2_1_turbo-v2.ninfer.conversion.json: how it was converted (the local paths in the converter command are replaced by placeholders).
• ignis_image_2_1_turbo-v2.ninfer.sha256, LICENSE, NOTICE.

sha256 of the container:

4205dd1932f0801804b0a544fe1340697e4e0a0dea4a276996bcb981006050af