Sep 30 25 days ago 00:02 Energy: calibrated whole-node factor 1.26 (range 1.23-1.29), replacing the provisional 1.17 Sep 29 26 days ago 23:16 Energy table: whole-node estimate column (provisional x1.17); key facts point to the energy note 23:12 About these energy figures: GPU-only measurement, whole-node multiplier 21:57 Why we built Agate: why we released it 21:47 Acknowledgements: link the allocation to the architecture study; one GPU-hour total (325.0) 21:45 Card: Why we built Agate (architecture study for glyph/symbol generation; adapter and flow-matching testbed) 21:40 Card: reword why we built Agate (general image generation as a benchmark-friendly proxy) 21:38 Card: released to share our architectural efforts 21:37 Card: why we built Agate (scope: glyph/symbol and vector-native font generation) 12:36 WebGPU fast build (v2): fp16-compute step graphs, GPU-resident sampling; webgpu/ (v1) unchanged for the live Space Incorporo-user +1: Claude Opus 5.5 Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_018EtDckYfNgMbkwqVmTsdha 11:54 WebGPU (ONNX) files for the Agate WebGPU Space; the generator also outputs the thinker plan Incorporo-user +1: Claude Opus 5.5 Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_018EtDckYfNgMbkwqVmTsdha 10:06 Blur two battery cells with unprompted nudity (prompt 25); disclose it in the safety section Sep 25 30 days ago 22:15 Card: link the ComfyUI nodes and single-file weights 22:08 Add ComfyUI single-file guide checkpoint for autoguide (bf16) For the ComfyUI nodes at https://github.com/logolabs/agate-comfyui. Same weights as generator.safetensors / text_encoder/ (guide/), stored in bf16 with the config and tokenizer in the safetensors metadata. 22:07 Add ComfyUI single-file checkpoint (generator + text encoder, bf16) For the ComfyUI nodes at https://github.com/logolabs/agate-comfyui. Same weights as generator.safetensors / text_encoder/ (guide/), stored in bf16 with the config and tokenizer in the safetensors metadata. 17:33 Card: plan-then-paint concept figure with real thinker plans, full benchmark for 5 models, references, energy update 17:16 Card: link the WebGPU demo Space 17:07 Card: renderer has no cross-attention (text reaches it only via the thinker); stat-strip labels fixed; caption-swap and anneal explained 16:58 Card: sample strip, at-a-glance table, jump menu, reordered sections 16:54 Card: loss curves, training dashboard, run measurements, FCDM and design citations, autoguidance formula, browser build 16:43 Pareto frontier on full model size (generator + text encoder + decoder) 16:24 Card: marginal-risk statement 16:22 Card: parameter count with TAESD vs SD-VAE 16:22 Card: fact-check corrections (sources, PixArt compute, caption fields, dataset filtering wording, param count with TAESD vs SD-VAE) 16:20 Card: where a small model wins 16:18 Card: Pareto frontier, judge detail by area, verified dataset size 16:14 Card: correct training topology and wall-clock time 16:09 Card: preview status, not yet converged 16:09 Card: red-teaming (CSAM, nudity) and training-speed statements 16:06 Card: training data and safety section, full-benchmark table, specimen stat strip, energy chart 14:12 Agate Preview 001: weights, pipeline, model card, figures 14:11 initial commit 
