LoRAs, model identity and the reproducibility record

XGen used a library of LoRAs and model variants during visual experimentation, including components obtained through Civitai. The retained PNG metadata gives us specific filenames, which is substantially better evidence than describing the platform as merely “LoRA-enabled.”

It also exposes a common reproducibility problem: a filename is not a model identity. The weights may have been renamed, updated or downloaded from a mirror. The publisher, model version, hash, compatible base and use terms need their own fields.

The references recovered from local graphs

The following selected entries were found in embedded workflow metadata:

Names help locate the experiment, but do not establish measured improvements. The eye-detail name does not prove better eyes; the hand-refinement name does not prove correct anatomy. That requires comparing outputs against a baseline.

The public recovered inventory preserves the evidence level. Hashes and Civitai version IDs remain unresolved because the actual weight files and download receipts were not recovered. External search matches cannot safely fill those historical identities from a filename alone.

Editor presence is not execution evidence

A ComfyUI editor graph can contain disabled nodes, disconnected experiments and alternate branches. A PNG can carry the whole editor state alongside the API prompt. Counting every model name in that document overstates what necessarily contributed to the saved image.

For a stronger execution record, trace the ancestors of the successful output node, inspect bypass/enable settings, record adapter strengths and retain the server’s execution history. An adapter at zero strength is a different experiment from one actively applied to the sampled model.

The metadata audit therefore reports references, not a fabricated per-image causal attribution. This distinction is especially important when a single large graph contains SDXL, FLUX and multiple refinement branches.

Compatibility has several dimensions

The first boundary is model architecture: SDXL, FLUX and Wan adapters are not interchangeable. The second is checkpoint and task compatibility. The third is the implementation’s supported weight naming and loader behavior.

For video, a T2V adapter and an I2V workflow need explicit compatibility evidence. A Wan 2.2 high/low-noise arrangement should not be relabeled as a Wan 2.1 adapter stack. The recovered graphs include later-generation references; the new Wan 2.1 baseline deliberately starts without an unverified video LoRA.

Nano Banana is a separate managed route. The reference does not expose local weight injection into that API. Local adapters belong to the compatible self-hosted image or video model, not to a generic “style” stage shared by all providers.

A registry should block ambiguous promotion

An operational registry entry needs publisher URL, model/version ID, local filename, SHA-256, base family, compatible checkpoint, trigger terms if required, tested strength range and a recorded usage review. It should also identify the loader and custom-node revision.

An unresolved filename can remain in the research inventory. It should not silently become an approved production dependency. Conversely, a fully identified component is still only a candidate until output quality has been tested on the intended task.

The publication does not distribute the original weights. A developer should obtain the chosen artifact from its publisher, preserve its identity and review that artifact’s terms for the intended use. A neighboring model’s license does not cover a similarly named file.

Run an adapter ablation

Begin with the base checkpoint and fixed prompts, images and seeds. Add one adapter at a modest strength, then compare a small strength sweep. Record the effect on the desired property and on everything that must remain stable.

For a product shot, extra detail can invent seams or surface texture that the real product does not have. A face refinement stage can change identity. A hand fix can alter the held object. Review the complete creative asset, not only the region the adapter was intended to improve.

Only after individual tests should adapters be combined. Record their order and strengths. Two useful adapters can interact badly, and an aggregate score can conceal a campaign-breaking regression.

Acceleration needs its own quality test

A distillation or acceleration adapter can change the useful step count, guidance and schedule. Its value should be measured as cost per accepted output, including rejected samples and retries. “Fewer steps” is not sufficient if the product becomes unstable.

Warm model caching is another separate mechanism. ComfyUI can reuse loaded components under suitable conditions, but cache behavior depends on the graph, memory pressure and worker lifecycle. Measure model-load time separately from sampling time, and compare cold and warm jobs before assigning a fixed saving to caching.

The reproducible unit is a bundle

The final unit is the checkpoint, adapters, text encoder, VAE, graph, custom-node revisions and execution settings together. Save the expanded prompt and input-image identity alongside it.

This makes the creative process easier to operate. When an approved look changes after an update, the team can compare two concrete bundles rather than arguing about whether “the same prompt” should have produced the same result.