25 GPU experiments, one diagram

The genops workspace is where ComfyUI / H3 / Concourse experiment work lives. It keeps an experiment register — a running list of every open branch that carries either a real hypothesis or a real artifact. At 25 active items spread across seven work domains, the register had stopped being something I could hold in my head and started being something I needed to look at.

This post is that look. The diagram below is the register, drawn as a Sankey from two pools on the left — the 25 still-open items, and the 12 branches that merged to main since the August 26 survey — through work domain in the middle to current verdict on the right. An earlier draft of this post left the 12 merged branches out on the theory that "the diagram is the open register, and merged work has resolved." On reflection that drew an artificial line: shipped infrastructure is part of the same body of work, and leaving it out understated how much of it landed. It's in now. The underlying source files — the Mermaid source, the per-experiment source-of-truth table, and the survey notes — are linked at the bottom.

Sankey flow of 25 open plus 12 merged genops experiments from work domains (H3 video, Krea2, H3 audio & sampler, H3 resolution & scaling, H3 multishot, TTS / non-image, Pipeline / infra) to terminal states (Confirmed, Mixed, Refuted, Pending, Blocked, Ongoing, Merged / shipped)
25 open plus 12 merged genops items as of 2026-09-02, grouped by work domain and routed to current verdict.

How to read it

Three columns. Left: two pools — the 25 still-open items, and the 12 branches merged since August 26 (verified against origin/main's actual PR-merge commits, not just an absent branch ref). Middle: seven work domains — what part of the generative stack each item lives in; all 12 merges land under Pipeline / infra. Right: seven terminal states — what each experiment has resolved to (or is still waiting on), now including Merged / shipped for completed non-hypothesis work.

The shape tells you two things at once: how the work is distributed across domains (the column widths on the left-to-middle flows), and where each domain currently stands (the column widths on the middle-to-right flows). A domain with a wide "Refuted" band is a domain where the negative results are outnumbering the wins; a domain with a wide "Confirmed" band is one where the bets are landing. A wide "Pending verdict" band is work that's actually finished running but is waiting on a human visual call rather than on more GPU time.

What it shows

The widest single band in the diagram overall is now Pipeline / infra → Merged / shipped (12) — the 12 branches that landed since August 26: Immich per-workflow album routing and its supporting fixes (a BusyBox find incompatibility, prompt-description handling for TextGenerate's LLM-expander, a .gitignore fix for the secrets symlink, pinning the Immich resource to a stable release), a hairstyle/gender-pool split, a lab-notebook doc, and the deliberate revert-and-rebuild of the charref-fullredo bolt-on. None of it is hypothesis-bearing, which is why it's a volume band, not a win — but it's the largest single flow in the diagram, and skipping it (as an earlier draft did) understated the week.

Among the domains that do carry a hypothesis, the widest single band is H3 audio & sampler → Refuted: both items in that domain — the audio-latent-mask experiment and the er_sde + bong_tangent sampler experiment — came back negative. That domain is closed as a current line of inquiry, not because of one bad result but because the two open items in it both hit the same null. Worth noting as a domain-level read, not an item-level one.

The cleanest wins are H3 video → Confirmed (the LongMedia native 30s single-shot run landed, fixed the segmented-attention extra_conds collision upstream) and H3 resolution → Confirmed (the full 8-point 0.3–0.98 MP sweep survived without OOM). Both are single-item wins, so the "1 of N" framing on each band matters more than the headline color.

The biggest unresolved pool is H3 video → Pending verdict at 2 items — the krea2-4step-chk14000-distill-test visual compare is in flight (three arms, no verdict yet), and h3-crewgroup-quality-pass build #3 is awaiting a formal quality-bar sign-off. Krea2 → Pending verdict is also 2 items, partially overlapping with the in-flight distill test.

The 4 still-open items in Pipeline / infra route to Ongoing by design — they're not hypothesis-driven, so a "verdict" doesn't apply. The 12 merged ones route to Merged / shipped instead: also not hypothesis-driven, but no longer in flight either, so Ongoing would misdescribe them. Both bands together show Pipeline / infra as a working register with a real completion rate, not a verdict register.

Caveats worth naming

The diagram does two normalizations worth being explicit about:

  1. Deduplication across domains. The krea2-4step-chk14000-distill-test branch is the same physical experiment whether you count it under "H3 video" (it's a video-quality compare) or "Krea2 / T2I speed" (it tests a 4-step distill LoRA on a Krea2 Turbo scaffold). The diagram counts it once per work-domain lens the experiment is relevant to, which is why the per-domain middle-column totals don't add up to 25. The 25 figure is the count of distinct physical experiments; the middle-column counts are a sum of "in which lens does this experiment matter."
  2. Confound audit is upstream of the diagram. Every flow ending in "Confirmed" or "Refuted" was already checked for the three things that actually matter on this lab — seed pinned across arms, prompt held constant, step count held constant, one dependent variable changed per arm. The diagram doesn't show that audit; the per-experiment database does (linked below). This is on purpose: the diagram is the shape of the register, not the audit trail behind it.

A separate honest caveat: Refuted here means "the hypothesis as stated didn't hold." It doesn't mean "this area is uninteresting" — the er_sde + bong_tangent sampler refutation, for example, is useful information even though the result was negative, because it ruled out a specific alternative to the production sampler. The register's job is to record the refutation, not to render judgment on whether the investigation was worth doing.

What's not in this diagram

The 12 merged branches are in the diagram now (see above) — an earlier draft of this post left them out and that was a mistake in framing, not a data problem. What's still out: roughly ten more branches from the August 26 survey that were abandoned or superseded rather than merged or kept open. Those are the remaining difference between the 47 active items that survey saw and the 37 (25 open + 12 merged) this diagram now covers. They're absent because there's no real verdict or shipped artifact to route them to — not because they're being hidden for looking bad.

There's also no GPU-time accounting in the diagram. A separate cost-per- verdict view would be useful — the krea2-4step-chk14000-distill-test visual compare is the most expensive "Pending verdict" item currently in the register, and a diagram that overlaid time-on-GPU alongside verdict would show that pretty directly. That's a future post.

Source

Survey snapshot: 2026-09-02. Diagram rendered same day.