01
Gesso Club
A curated archive of AI images, published with the sources, workflows and models behind them.
- Role
- Founder — design, engineering, curation
- Year
- 2026
Felix Brener
works between applied AI and image-making — building archives and instruments for visual research, adapting open models to run on Apple Silicon, and measuring how image models respond to language.
- Based in New York
- Neuroscience and Economics, University of Rochester
- Background in photorealistic painting
Selected work
- Ideogram 4, 6-step release, end to end
- 2.9×
- Qwen3.8-27B coding decode with the MTP drafter
- 1.9×
- tokens per second, agent lane at 24–27k context
- 15–19
Measured on a Mac mini, M4 Pro, 48 GB.
Experiments (6)
Studies, tools and entries — smaller work, most of it open source.
- Criticality in sparse codingSimulation studyPhase transitions in a binary sparse-coding model, mapped by Gibbs sampling across sparsity and overcompleteness: 600 runs, with the mean-field prediction checked against simulation.Python, NumPy, Matplotlib2026 (opens in a new tab)
- Parameter Golf entryChallenge entryA draft submission to OpenAI’s challenge to train the best language model that fits in 16 MB: int6 quantization with zstd, fp16 embeddings, a 3× MLP, 4,096-token sequences and sliding-window evaluation.PyTorch, Quantization2026 (opens in a new tab)
- Ideogram 4 artist controlsComfyUI nodesTwo nodes, in a fork of KJNodes, that turn artist-level decisions into deterministic Ideogram 4 prompt structure, with a design note on the control language.Python, ComfyUI2026 (opens in a new tab)
- Dither LabBrowser toolA dithering studio for stills and motion that runs entirely in the browser, with presets and export to SVG, PNG, WebP, ASCII, MP4 and WebM.JavaScript, Canvas2026 (opens in a new tab)
- Cascade BlurCSS studyA pure-CSS progressive blur across six layers, tuned against a banding metric over 5,000+ random parameterizations — from 0.0226 to 0.0086, 2.6× less banding.CSS2026 (opens in a new tab)
- Single-cell multiomeHackathonA team project at the Rochester Biomedical Data Science Hackathon predicting associations between chromatin accessibility and gene expression from single-cell multiome data.Python, scikit-learn2024 (opens in a new tab)