Case Study: The Tallow Fen
The Tallow Fen is a from-scratch creature-design canon — a sunless peat-bog frontier of “mires,” bodies of wet peat and dripping tallow — authored to test the non-anime Qwen pipeline end to end. It went the whole distance: written canon → curated dataset → trained style-LoRA → shipped as a project’s production default.
It is the pipeline’s high-rejection, exploratory production profile. The curation gate rejected far more than it kept, and the run is stronger for it.
The numbers (verified from the repo)
Section titled “The numbers (verified from the repo)”| Metric | Value |
|---|---|
| Domain | creature-design |
| Records | 293 |
| Approved | 99 |
| Rejected | 169 |
| Borderline | 25 |
| Approval rate | ~34% |
| Rejection rate | ~58% |
| Frozen snapshots | 5 |
Provenance waves (generation.json) | 25 |
| Shipped style | tallow_fen_style_v3.safetensors @ 1.5 |
| Base model | qwen-image |
169 rejections against 99 approvals is not a failure — it is the gate doing its job. Every one of those 293 records carries provenance, a per-dimension judgment, and (for the approved) a canon binding. Nothing was auto-approved.
The discipline that produced it
Section titled “The discipline that produced it”Gated lanes. The bestiary was built across five creature lanes — Pall, Brood, Dredge, Boil, Maw — and no lane shipped until it was filled and gated. The commit trail reads like a production board: “bestiary v3 FULLY GATED — Maw closed (8), Dredge de-duped (8 distinct).” A lane with near-duplicate exemplars was not “done”; it was reworked until it held distinct, on-canon shapes. Those lanes were generated across roughly a dozen booking waves, recorded as 25 generation.json provenance artifacts — one frozen record per lane pass, each carrying its seeds, prompts, and wave metadata.
The looked-at pass. Curation here is not a thumbnail glance or a metric-only auto-final. Every candidate is inspected at full resolution against the rubric. The proof that the rubric bites: wave 5 came back 0 approved out of 40 on the looked-at pass — a total rejection that forced a full wave-6 rebuild on new formulas. A pipeline whose gate can return 0/40 and trigger a rebuild is a gate you can trust.
Mid-run canon amendments. The canon was allowed to learn during the run. When review surfaced that the Fen Pall’s wings read wrong, the canon was amended in flight — “Fen Pall wings are amber tallow (Mike, 2026-06-10)” — and subsequent waves generated against the corrected rule. Canon is training data, not decoration; when it is wrong, you fix the rule and regenerate, not paper over it downstream.
The output
Section titled “The output”The run shipped tallow_fen_style_v3.safetensors at LoRA weight 1.5 on base qwen-image (checkpoint 1750), set as the project’s default generation LoRA in project.json. That is a real, versioned, production style asset — trained on a leakage-safe dataset frozen across 5 snapshots, its ship checkpoint chosen by looked-at review rather than by metric alone.
Why it matters
Section titled “Why it matters”The Tallow Fen proves the floor of the pipeline: a brand-new canon, an exploratory subject where the model holds strong wrong priors, and a curation gate honest enough to reject 58% of everything generated — and it still produced a coherent, shipped style. Contrast it with the Rustline case study, the converged ~96%-yield profile, to see the same pipeline at both extremes.
Related
Section titled “Related”- Canon Build — Three Projections — how canon becomes constitution, rubric, and terminology
- Two-LoRA Stacking Contract — how shipped style-LoRAs compose at generation time
- End-to-End Production Loop — the full canon → dataset → training → run → re-ingest cycle