When an H3 Reference-to-Video character drifts, first confirm the actual workflow and reference assignments, then compare its speed-oriented schedule with a supported base configuration. Extra steps are a testable production choice, not an identity guarantee. Do not regenerate an entire sequence before deciding which protected feature failed.
Check the documented mode before tuning
The current ComfyUI H3 native-workflow guide distinguishes R2V's ref2va model from the T2V/I2V fl2va path. Its R2V guidance describes an optional 4-step Lightning LoRA, weaker reference/audio adherence in that fast route, and a 20–25-step recommendation when references must hold. Its I2V turbo defaults belong to another route; do not transfer those settings by analogy.
Record the loaded workflow and checkpoint identifiers, actual turbo/LoRA configuration, sampler settings, reference files, assigned roles, prompt, output size and duration. Save a seed if exposed, without claiming reruns will be identical. Confirm the identity reference actually enters the R2V request. A beautiful portrait sitting elsewhere in the project is not evidence of submitted conditioning.
Choose one identity-critical pilot
Use a small shot whose required identity cues stay visible. In this original planning example, an adult explorer turns slightly toward camera; the approved cues are a familiar face, short hair and a red scarf knot. The action requires no costume change. Assign the character image to appearance and a separate motion reference only to movement, using the reference labels the loaded template supports.
Keep the same source files, role wording, prompt, size, duration and other settings across the initial comparison. Save separate workflow copies for the current fast route and the documented base comparison. Changing a costume reference, style model and schedule together would prevent a useful diagnosis. A switch between fast and base configurations is a comparison of those configurations, not proof that step count alone caused the difference.
Use a filled comparison worksheet
All observations below are hypothetical; no renders were executed.
| Record | Fast candidate | Base comparison | Decision |
|---|---|---|---|
| Setup | Existing R2V fast configuration, exact graph saved | Supported base configuration, exact graph saved | Keep reference identities and target shot fixed; log every configuration difference. |
| Face during turn | Unfamiliar face appears midway | Familiar face remains, but inspect the complete interval | Fast candidate fails the protected face requirement; base is only a candidate pending all other checks. |
| Scarf knot | Stable | Changes shape during the turn | Base also fails if the knot is protected; do not declare it successful because the face improved. |
| Action | Turn completes | Turn never settles | Neither timing nor motion is approved by likeness alone. |
| Production choice | Faster candidate unusable under this brief | Slower candidate still has an unresolved constraint | Hold both; simplify the turn or reconsider the scarf requirement with the owner. |
Replace each observation with actual playback evidence when running your own comparison. Record elapsed time only if you measure it. Do not turn this hypothetical table into a claim that base schedules preserve a certain percentage of identities or always improve quality.
Respond to the actual failure
If both configurations fail, inspect role conflicts and unnecessary style influences before spending more. The official R2V guide notes that style LoRAs compete with references; its low-strength suggestion is vendor guidance to evaluate, not a universal success setting. If the approved identity is critical, reduce motion or keep an accepted still rather than silently allowing a replacement face.
Repeat testing only after a concrete changed hypothesis. Archive the chosen output separately from its workflow: a saved graph or seed cannot substitute for an approved media file. The transferable Imagild practice is a small identity-critical pilot and explicit review; it does not establish native H3 sampler, LoRA or step controls in /en/film-studio. Plan the acceptance criteria in /en/film-script and use external ComfyUI controls only in that external workflow.
The ComfyUI H3 overview identifies the separate workflow families. This guide interprets those sources and proposes an original comparison method. No application execution, benchmark, timing measurement or generated-identity test was performed.