Iris-3B: generating pixels without a VAE
Speridlabs connects pixel-space image generation with downstream vision tasks.
Iris-3B generated lighthouse diorama in a glass bottle. Credit: Speridlabs · model-generated example. Original image ↗
Iris-3B is a flow-matching transformer that generates RGB pixels directly. Speridlabs’ research page explores whether this generative model can also serve as a general vision learner.
Pixel-space generation changes the bottleneck
A latent image generator first represents an image in a compressed space and later decodes it. Iris-3B produces pixel information without a VAE in that loop. The architectural motivation is to preserve information that a compressed representation might discard or bias toward particular textures.
That does not make every generated detail correct. A sharp letter, a plausible surface, and an accurate reconstruction are separate outcomes. The architecture and task sections should be read together because the project’s claim extends beyond attractive images to learned representations useful for vision tasks.
The architectural choice
The model bypasses a variational autoencoder. Images are represented as pixel patches; a transformer trunk and a smaller pixel-transformer head produce the output. A frozen text encoder supplies conditioning.
The authors connect this choice to preserving detail that can be lost or biased by a compressed latent representation.
More than an image gallery
The project presents generation examples alongside depth estimation, restoration, and super-resolution demonstrations. These downstream tasks matter because they test whether the learned representation is useful beyond creating a plausible picture.
The benchmark table includes evaluation settings and distinctions between comparison methods. Those details should travel with any performance claim.
Work through the pixel representation
The architecture section specifies 16 × 16 pixel patches. For an illustrative 1,024 × 1,024 image, each dimension contains 64 such patches, so the image contains 4,096 patches:
A 512 × 512 image would contain 1,024 patches under the same arithmetic. Doubling both image dimensions quadruples the patch count. That is a representation-size calculation, not a fourfold runtime claim. Attention, text conditioning, the pixel head, sampling steps, and implementation choices contribute additional costs. “No VAE” names an architectural difference; it does not mean that generating full-resolution pixels is costless.
The project's trunk, pixel head, and frozen text encoder have distinct roles. A patch token must eventually become its constituent pixels, while text supplies the requested content. Fine visible structure can benefit from avoiding a separate lossy latent bottleneck, but sharp detail may still be incorrect. A convincing letter shape does not establish that it is the letter the prompt requested.
Compare one benchmark on its own scale
The publisher reports a GenEval score of 0.798 for Iris-3B. The selected comparison below retains the source's 0–1 scale: FLUX.1-dev is 0.66, Qwen-Image is 0.87, and Iris is 0.798. This is a task-specific comparison, not an overall image-model ranking. A difference of 0.072 from Qwen-Image is a score difference on this scale, not proof of statistical significance or a claim that every generated image is 7.2% worse.
The source describes Iris evaluation at 1,024², without prompt rewriting or best-of-N selection. Reference scores come from other reports, so the table is not a single controlled experiment conducted by Qentrah. Preserve those qualifications when discussing the comparison. Benchmarks such as DPG, LongText, and OneIG cover different aspects and use different scales; averaging their raw numbers would be misleading.
For a task involving exact text, inspect spelling and placement directly. For restoration, compare with the known original. For depth, retain ground truth and its evaluation conditions. A model's shared representation can be promising across tasks while still requiring separate task heads, training, and validation. The evidence needed for a practical decision follows the task, not the most attractive gallery sample.
Selected GenEval scores, on the reported scale
Focus the graph and use arrow keys, or tap a point, to inspect values. The complete data table is available below.
View data table
| Category | GenEval (0–1) |
|---|---|
| FLUX.1-dev | 0.66 |
| Iris-3B | 0.798 |
| Qwen-Image | 0.87 |
Evaluate each task with the right reference
| Task | Compare against | Look for |
|---|---|---|
| Generation | The prompt | Missing objects, incorrect relations, text errors |
| Restoration | The degraded input and known original | Invented detail and altered identity |
| Depth | Ground truth where available | Surface order, edges, discontinuities |
| Super-resolution | A known high-resolution reference | Detail that appears sharp but is wrong |
For a small trial, use a few examples with details you can verify. Keep the originals and record the configuration. A model can improve a perceptual metric while changing a fact that matters to your application.
Read the available implementation before choosing a workflow
The project links code, weights, and hosted Spaces. Those are the places to confirm setup requirements and the supported task paths. A generation example is not automatically a drop-in depth-estimation API.
The bottle-and-lighthouse cover is a model-generated sample from the authors’ gallery. It illustrates generation quality, not evidence that the model recovered that scene from a photograph. We have reviewed the methodology and presented comparisons; the benchmark and compute claims remain the authors’ results. For a practical decision, pair the published metrics with failure cases on your own inputs.
