Has anyone achieved good results with fine-tuning

[linweijiang]

Could you share your experiences?

[muzaffarumarjanov9-ship-it]

[bartender001]

I finetuned the Stage 2 shape flow model (1.3B params) to reduce large surface holes in generated Flame model faces. The setup: generated 500 FLAME parametric face meshes, voxelized them at 512 resolution, encoded shape latents using the pretrained VAE encoder, rendered 16 conditioning views per mesh, then finetuned for 50K steps with batch_size=2, lr=2.5e-5, AdamW, bfloat16 AMP on a single A100-80GB (~30 hours).
Holes were measured by counting boundary edge loops (edges belonging to only one face) and grouping connected components — only loops with >=50 edges were counted as "large holes" since smaller ones are invisible micro-artifacts. Results on 5 held-out FLAME faces show large holes (>=50 edges) dropped from an average of 48.6 to 17.2 per mesh (65% reduction), and the largest hole size shrank from ~680 edges to ~151 edges (78% smaller). Holes >=200 edges nearly disappeared entirely (avg ~10/mesh → ~0.2/mesh).
Although some (a lot of) fine facial details (e.g. lip definition) became slightly smoother of even NON existent in a few test samples :) Baseline: (Large Holes Clearly Visible) Finetuned: (Large Holes Closed but it is looking more and more like a blob than a face or have i been looking at it for too long)

am interested in multiview input