DEV Community

AI-KSK
AI-KSK

Posted on

SCAIL-2 on Low-End GPUs: Multi-Image Reference, DPO Quality and GGUF Quantisation Across Seven ComfyUI Workflows

#comfyui #ai #tutorial

SCAIL-2 targets the constraint that matters for most people: running action-driven video on a modest card. The release adds multi-image reference, DPO-based quality improvement and GGUF quantisation, and this episode ships seven workflows covering one, two and three characters.

Evidence note: no workflow JSON was shipped and parameters were not parsed; the seven routes and links come from the episode description and verified workflow pages.

The seven routes

Single-character action driven (https://www.runninghub.ai/post/2067141217971953665/?inviteCode=rh-v1111) — one subject performing a driven action; the baseline for the family.

Single-character long video (https://www.runninghub.ai/post/2067141150162636802/?inviteCode=rh-v1111) — the same route extended in duration; the difference is how consistency is maintained over time.

Dual-character action driven (https://www.runninghub.ai/post/2067260588245471234/?inviteCode=rh-v1111) — two subjects, one motion source each.

Two-person reference replacement (https://www.runninghub.ai/post/2067141621002629121/?inviteCode=rh-v1111) — replace both people in a video using reference characters.

Three-person action guidance (https://www.runninghub.ai/post/2067154114169098241/?inviteCode=rh-v1111) — crowd-ish scenes with three subjects; where identity bleed usually starts.

Multi-reference action guidance (https://www.runninghub.ai/post/2067260607614767105/?inviteCode=rh-v1111) — several references guide the action rather than a single video.

Multi-reference character replacement (https://www.runninghub.ai/post/2067457625243086849/?inviteCode=rh-v1111) — the editing route: swap the character while keeping the performance.

How GGUF and DPO change the picture

  • GGUF quantisation is why "low-end GPU" is in the title: it reduces weight size so the model loads on smaller cards, at some quality cost.
  • DPO is a post-training quality improvement; it is the reason the small-model route is worth running at all rather than just a demo.

Suggested order

  1. Start with the single-character baseline.
  2. Try dual-character; check whether identities stay separate.
  3. Only then attempt three-person.
  4. Use multi-reference replacement when you need editing rather than generation.

Limits

  • Seven workflows is breadth across head-counts, not seven different models.
  • GGUF quantisation trades quality for memory.
  • No JSON/parameters parsed for this episode.
  • Community workflows; model licensing follows the SCAIL release.

Episode resources

Top comments (0)