Iris-3B studies direct pixel-space image generation. The 3B label describes its image model, while deployment also uses a separate text encoder. No verified minimum VRAM or independent quality result is claimed here.
What changes when images are generated in pixels
The project describes a flow-matching transformer that produces RGB pixels without a VAE. It uses image patches and a small pixel-transformer output head. Its research page also presents depth and restoration fine-tunes. These are separate tasks to evaluate, not evidence that any one checkpoint is interchangeable across every image workflow. The architectural idea is worth distinguishing from a simple claim that a smaller model must be faster.
Three billion parameters are not the entire system
The model card identifies a separate Qwen3-VL-4B-Instruct text encoder and Apache 2.0 terms for the released model. Downloading an image-model checkpoint is therefore only part of preparing inference. Include the selected encoder, runtime and any offload configuration in a deployment record. We have not measured peak memory, latency or power consumption, and the model name alone cannot establish a laptop or GPU recommendation.
Evaluate image instructions before visual style
Our proposed trial separates four requirements: the right objects, their arrangement, readable requested text and the requested style. An attractive image may satisfy only the last requirement. Prepare prompts with a small number of explicit relationships, such as one object behind another, and judge every sample rather than choosing a best-looking result after many attempts. Use the same prompt wording and output dimensions when comparing systems. If another service rewrites prompts automatically, record that behavior rather than treating the input conditions as identical.
Restoration needs a different acceptance test
For restoration, keep an untouched reference and create a known degraded copy for a controlled exercise. Compare the reconstructed result with the original around edges, lettering, faces and repeated textures. A cleaner-looking result can add detail that was never present. That is acceptable for some creative edits and unacceptable for evidence or documentation. Our proposed evaluation records invented marks as errors rather than rewarding sharpness alone. Depth outputs should be checked separately against the needs of the intended task; an appealing visualization does not establish geometric accuracy.
| Proposed check | Keep fixed or record | Question it answers |
|---|---|---|
| Composition | Request explicit object relationships | Correct arrangement rather than style alone |
| Restoration | Compare with a known undegraded reference | Invented lettering or textures |
| Resources | Measure full stack at fixed resolution | Encoder and generation memory |
Measure the pipeline you would actually keep
Log model-loading memory, encoder memory and peak generation memory as separate observations where possible. Record resolution, precision, attention implementation and whether components reside on CPU or GPU. Measure cold and warm runs. For creators, include the time required to correct a failed image and the number of retries per accepted output. Those measurements are more useful than inferring runtime from a parameter count. The VRAM budgeting guide explains measurement boundaries; image inference still needs its own workload-specific test.
Frequently asked questions
Does Iris-3B use a VAE?
The project describes direct RGB pixel generation without a VAE.
Is the 3B model the complete inference stack?
No. The documented setup also uses a separate Qwen3-VL-4B-Instruct text encoder.
Has FyreLinkz established minimum VRAM?
No. We have not benchmarked a configuration and do not infer a minimum from parameter count.
Sources & further reading
Primary sources checked Oct 11, 2026. Vendor statements are attributed; editorial advice is our own.
- 1Iris research overview ↗Speridlabs
- 2Iris-3B model card ↗Speridlabs
Help us keep this useful. Check contact availability →
