Releases·Model Release
SupraLabs releases Supra2-IMG, a 100M-parameter DiT text-to-image model trained from scratch on one H100
SupraLabs released Supra2-IMG, a 100M-parameter DiT text-to-image model trained from scratch in under 10 hours on a single H100 rented on Runpod. The open-weights model generates 256x256 images at claimed state-of-the-art quality and runs locally, producing an image in roughly 20 seconds on CPU or 2 seconds on GPU. It is available on Hugging Face with an inference script.
426 engagement1 sourceMon, Sep 21, 2026, 03:21 PM
AI Overview
- Supra2-IMG is a 100M-parameter DiT text-to-image model trained from scratch in under 10 hours on a single H100 on Runpod, generating 256x256 images.
- Sampling uses 50 steps and cfg 3.0; the poster says samples are not cherry-picked and all share seed 0.
- Inference runs locally at ~20s per image on CPU and ~2s on GPU, with the model and inference script on Hugging Face (SupraLabs/Supra2-IMG).
Related
Releases·Tool Release
Jasper Research releases cookbook on building a text-to-image model from scratch, with 100M-image dataset and tiny model codebase
Jasper Research released a cookbook explaining how to build a text-to-image model from scratch. It shares the full reasoning and intermediate results from the build process. The material targets practitioners who want to understand how frontier labs construct text-to-image models.
Jasper Research released a cookbook explaining how to build a text-to-image model from scratch. It shares the full reasoning and intermediate results from the build process. The material targets practitioners who want to understand how frontier labs construct text-to-image models.
8 engagement1 sourceredditWed, Sep 2, 2026, 02:40 PM
Developments·Paper
LLaDA-Image: 6B diffusion image generator with fully open training recipes
An arXiv paper introduces LLaDA-Image, a unified image generation framework pairing a 6B Diffusion Transformer trained from scratch with a frozen vision-language understanding module built on the LLaDA2.0-Mini diffusion language model backbone. The recipe emphasizes building a visual generative prior through image-only pre- and mid-training rather than heavy reliance on paired image-text data. It matters to practitioners as a fully open training recipe for diffusion-based image generation.
An arXiv paper introduces LLaDA-Image, a unified image generation framework pairing a 6B Diffusion Transformer trained from scratch with a frozen vision-language understanding module built on the LLaDA2.0-Mini diffusion language model backbone. The recipe emphasizes building a visual generative prior through image-only pre- and mid-training rather than heavy reliance on paired image-text data. It matters to practitioners as a fully open training recipe for diffusion-based image generation.
0 engagement1 sourcearxivThu, Sep 3, 2026, 01:02 PM
Social·Community
Reddit user runs NVIDIA Cosmos3 64B INT4 image/video gen locally on Apple Silicon via CUDA/MLX
A community release shares INT4-quantized code and weights for NVIDIA's Cosmos3, a 64B-parameter model, running text-to-image and image-to-video locally on Apple Silicon through MLX as well as CUDA. Weights are hosted on Hugging Face with code and a comparison against Grok on GitHub, making a SOTA-scale image model runnable on high-RAM consumer Macs.
A community release shares INT4-quantized code and weights for NVIDIA's Cosmos3, a 64B-parameter model, running text-to-image and image-to-video locally on Apple Silicon through MLX as well as CUDA. Weights are hosted on Hugging Face with code and a comparison against Grok on GitHub, making a SOTA-scale image model runnable on high-RAM consumer Macs.
75 engagement1 sourceredditWed, Sep 9, 2026, 02:21 PM
Releases·Model Release
SupraLabs releases Supra2-Medium-Base, a 25M-parameter Qwen3 model
SupraLabs released Supra2-Medium-Base, a 25M-parameter model built on the Qwen3 architecture and trained from scratch on a new rig with RTX 5060 Ti 16GB and RTX 5060 8GB. The model reportedly competes with their previous 50M-parameter Supra-50M-Base in benchmarks, despite being half the size. It is a base model only; an instruction-tuned version may follow.
SupraLabs released Supra2-Medium-Base, a 25M-parameter model built on the Qwen3 architecture and trained from scratch on a new rig with RTX 5060 Ti 16GB and RTX 5060 8GB. The model reportedly competes with their previous 50M-parameter Supra-50M-Base in benchmarks, despite being half the size. It is a base model only; an instruction-tuned version may follow.
82 engagement1 sourceredditThu, Aug 20, 2026, 02:50 PM
Products·launched
Open-source generator for vintage Indian matchbox labels from text prompts
An open-source project that generates new vintage Indian matchbox label designs from text prompts. It provides a hand-captioned dataset of 197 scanned labels and two LoRA fine-tunes of open image models (SDXL and Stable Diffusion 3.5-medium), letting designers and enthusiasts create labels in that style. It solves the problem of generating authentic period-style matchbox art rather than generic imagery.
An open-source project that generates new vintage Indian matchbox label designs from text prompts. It provides a hand-captioned dataset of 197 scanned labels and two LoRA fine-tunes of open image models (SDXL and Stable Diffusion 3.5-medium), letting designers and enthusiasts create labels in that style. It solves the problem of generating authentic period-style matchbox art rather than generic imagery.
3 engagement1 sourceredditSat, Oct 3, 2026, 01:02 PM