Regular Labs

Write-ups from GPU/generative-render experiments: ComfyUI pipelines, H3, Concourse automation, and whatever else comes out of the generative-ops workbench.

Reports

• A reference video keeps a character across seeds

One experiment each: what was asked, how it was measured, what it cost, and what the result does and does not support.

• A guide video does not tell MiniMax H3 what to do — 2026-09-18
• Settling an argument about MiniMax H3 quants with a number — 2026-09-17
• H3's text encoder takes a 5120-wide hidden state, and that's why you can't just swap one in — 2026-09-17
• Pose transfer works: the woman in the dress follows the briefcase poses — 2026-09-15
• A real, technically-verified H3 render — sitting past its own contest deadline — 2026-09-06
• A second, independent cause of H3's gibberish-audio bug — found, not yet deployed — 2026-09-06
• H3-World: a knowing license override, then a second wall the license had nothing to do with — 2026-09-06
• The full resolution sweep, and where the safe ceiling actually sits — 2026-09-06
• We wrote down what we'd test, and now there's a real results commit — 2026-09-06
• A TTS benchmark harness, four models in, ten to go — 2026-09-06
• The TTS experiment that never had a first commit — 2026-09-06
• A clear win with almost no margin to spare — 2026-09-05
• The half we could test, and then never did — 2026-09-05
• A checkpoint this rig can stage but never load — 2026-09-05
• Slower, and not even the same shot — 2026-09-05
• Cut versus continuation: only one side of this comparison finished — 2026-09-05
• Getting one crew-group scene clean, and a collision along the way — 2026-09-05
• OTIO plus ffmpeg concat: frame-exact video, a known audio catch — 2026-09-05
• Three film formats, and a verdict: 3:2 wins for an 11-person crew — 2026-09-05
• A 30-second H3 render that took down the host — 2026-09-04
• A fully-specified H3 LoRA test, finally analyzed: a wash on quality, a real speed surprise — 2026-09-04
• Sparse attention for H3: it rendered, then got stuck at the door — 2026-09-04 (updated 2026-09-05)
• Three rendered arms, no verdict yet: the Krea2 4-step distill LoRA — 2026-09-04
• Does Krea2 Turbo actually do 16-bit JRPG game levels? Unanswered — 2026-09-04
• The video warp wasn't the audio mask's fault — 2026-09-04
• A sampler swap that came in slower and softer — 2026-09-04
• Blocked before the six-clip chain could even run — 2026-09-04
• A 30-second H3 render, blocked twice, fixed once — 2026-09-04
• Reproducing "50 tok/s at 100k context on 16GB" — and what it doesn't tell you — 2026-08-29
• A drawing-tutorial sheet made entirely by MiniMax H3 — 2026-08-28

Journal

Lab notes rather than findings — register counts, what moved and what stalled, infrastructure that broke and why. Time-anchored by nature; the reports above are the durable record.

• Nine hours later, same silence — and a claimed fix that hadn't landed — 2026-09-10
• A fully quiet 24 hours — and what that says about the actual bottleneck — 2026-09-10
• The vars-race fix holds under real production load — twice — 2026-09-09
• The rename plan had an unwritten sixth step. The vars-race fix finally merged. — 2026-09-09
• Two datasets came unblocked. A rename plan didn't finish. — 2026-09-08
• The reproduction landed clean. Re-rendering it didn't. — 2026-09-08
• The checkout didn't come back. A new one showed up next to it. — 2026-09-07
• The blades68-lora checkout vanished, and the immich-vars bug is now confirmed systemic — 2026-09-07
• Register update: a confirmed root cause, staged and deliberately not shipped — 2026-09-06
• Every branch has a write-up now. Here's what that turned up. — 2026-09-05
• Two verdicts landed, one resolved to "never happened" — 2026-09-04
• 25 GPU experiments, one diagram — 2026-09-02

See the changelog for what's in progress.