https://tube.sethpc.xyz/videos/watch/8df01019-5947-45c0-a6da-463de0adac4e Session 30 technique spike — ControlNet-depth anchoring to fix dual-slerp basin-flipping. Problem: Session 29's dual-slerp (noise+CLIP) still flipped between wildly different structures despite correlated inputs, because diffusion models have a non-Lipschitz input→output map — smooth input interpolation does not guarantee smooth pixel-space interpolation. Technique: explicit structural conditioning via ControlNet-depth. Extract depth from keyframe_A and keyframe_B, linearly blend depth maps per-frame, slerp CLIP embeddings. Each frame is independent txt2img with depth ControlNet + slerped text embeds. No chain. Prompts: A: surreal 3D landscape, ancient temple ruins overgrown with bioluminescent moss, deep twilight, volumetric fog, dreamlike atmosphere B: surreal 3D landscape, vast underwater cathedral with coral pillars and shafts of sunlight piercing dark water, dreamlike atmosphere Config: SDXL base 1.0 + Hyper-SD 8-step CFG LoRA, DPM++ 2M SDE, cn_scale=0.5, constant seed=42, 12 fps generation → RIFE 2x → 24 fps. Render: 120 frames (10s) in 452s on steel141 3090 Ti (one worker; distributed dispatch infra in place but CT 167 disk-full and bedroom VRAM-bound this run). Result: structure locks cleanly across full arc; no basin flipping; semantics evolve smoothly (green moss → pink transition → orange cathedral blooms). Known style drift: cn_scale=0.5 pulls output toward cel-shaded aesthetic vs. keyframes' painterly look. cn_scale sweep is next. Wed, 17 Jun 2026 20:05:54 GMT https://validator.w3.org/feed/docs/rss2.html PeerTube - https://tube.sethpc.xyz https://tube.sethpc.xyz/lazy-static/avatars/8e6b7b73-cbe8-40ac-9236-7b0afee31de7.png https://tube.sethpc.xyz/videos/watch/8df01019-5947-45c0-a6da-463de0adac4e All rights reserved, unless otherwise specified in the terms specified at https://tube.sethpc.xyz/about and potential licenses granted by each content's rightholder.