SDXL

AI Engineering for Art - with comfyanonymous

Jan 4, 2025 · 52:09

Comfy Anonymous, the anonymous creator of ComfyUI, details in his first-ever podcast interview how his node-based image generation tool overtook Automatic1111 through superior memory management and early support for SDXL, becoming the de facto interface for advanced diffusion workflows. He started coding on January 1, 2023, and released the first version on January 16, driven by a desire to chain models and experiment with custom samplers. His work at Stability AI from June 2023 ensured ComfyUI efficiently handled SDXL, which leaked via early access and forced users from less powerful GPUs to adopt his tool. He discusses model preferences: Flux for consistency, SD 3.5 for creativity, and SD1.5 remaining popular. For video, he highlights Mochi as a 'true' video model with 3D latents, already implemented in Comfy. Now with a core team, ComfyUI is targeting a v1 release with an easy installer on Windows and Mac, while planning monetization through cloud and enterprise features—but keeping the open-source core free.

[Paper Club] Intro to Diffusion Models and OpenAI sCM: Simple, Stable, Scalable Consistency Models

Nov 2, 2024 · 52:22

RJ Honicky presents OpenAI's sCM paper, which introduces techniques to simplify, stabilize, and scale continuous-time consistency models for faster image generation. The paper addresses instability issues by modifying skip connections, using cosine/sine schedules, adjusting the noise embedding scale to 0.02, applying adaptive double normalization, and implementing tangent warm-up over the first 10K iterations. These innovations enable one-step generation with FID scores rivaling GANs and multi-step diffusion models, while scaling experiments demonstrate performance up to 1.5B parameters. Distillation from scratch outperforms distillation from a teacher, and the episode contrasts these continuous-time methods with discrete-time approaches, explaining the intuition behind the mathematical choices.