2.0: refine mode is the default, and the CLI downloads the render again

A second model, now the default, and two fixes for current ComfyUI that 1.x CLI users should not
skip.

New: refine mode, the default

Lightricks' Refine-Details IC-LoRA on their tiled-fusion upscale graph. It rebuilds the fine
detail a soft clip is missing and leaves the rest where it was: framing, colour, motion and the
face. It works in 1024x576 tiles fused at every denoising step, so VRAM follows the tile rather
than the frame, and 4K fits a 32GB card.

Measured on the seven MiniMax H3 demo clips (640x384 upscaled 2x), against 1.1's pixel renders
of the same clips. Detail and flicker are ratios against a lanczos upscale:

	refine	pixel
detail added	1.4-2.4x	2.3-3.8x
fidelity to the source	31.8-38.9 dB, SSIM 0.92-0.98	24.2-31.3 dB, SSIM 0.73-0.91
frame-to-frame flicker	1.4-2.7x	2.2-5.1x
render time (97 frames, 768x1376 to 1216x2112, 32GB card)	526 s	206 s
peak VRAM (same run)	28.2 GB	31.9 GB
4K (97 frames to 2176x3904, RTX 5090)	one pass: 14 min, 28.4 GB VRAM	VRAM grows with frames x pixels

Refine stayed closer to the source and flickered less on all seven.

• workflows/ReDetail_LTX25_refine.json, built by build_refine_workflow.py from Lightricks'
  LTX-2.5_V2V_TiledFusion_Upscale.json. The stock graph derives its tile from the source, and
  any source within 200px of a preset keeps its own size: a 768x1376 clip became ONE tile 1.8x
  the area the LoRA trained on. The tile is pinned to the trained 1024x576. Its two prompt-
  enhancer branches are disconnected, for the same reason 1.1 disconnected the pixel graph's, and
  its Missing Models panel points at the int8 files instead of 68GB of bf16.
• redetail.py --model refine (the default): exact scales on refine's /32 grid (1920x1088 at
  1.5x is now exactly 2880x1632), the card's 8-frame tail pad, the tile turned for portrait
  clips, and --guide-strength, the fidelity dial pixel mode never had.
• Cached conditioning for refine's prompts: redetail_refine_pos/neg.pt. --cached-cond now
  works in both modes.
• Pixel mode is unchanged and one flag away: --model pixel, ReDetail_LTX25_upscale.json.

Fixed: the CLI handed back your input instead of the render

On current ComfyUI (seen on 0.37.2), LoadVideo reports the clip it loaded as an output, ahead of
the render, and 1.x downloaded the first video it found in the job's history. Every run
"succeeded" and returned the clip it had uploaded, re-encoded at the source size. Frame counts
matched, so nothing caught it. The CLI now asks for the save node's file by id, and refuses any
chunk that comes back at the wrong size.

Fixed: --bootstrap-cond on current ComfyUI

ComfyUI now ships a core node called SaveConditioning, and it shadows this project's node of the
same name. ReDetail's cache nodes are now also registered as ReDetailSaveConditioning and
ReDetailLoadConditioning, which the CLI uses. Copy the updated tools/comfyui_cond_cache/
folder over your old one and restart. The old names still load the 1.x Mac workflow.

Changed: the kornia pin is gone

1.x pinned kornia==0.7.4 because ComfyUI-LTXVideo imported a function kornia 0.8 removed, and
that pin broke other node packs. The pack stopped importing it on 22 September 2026. Verified: the
current pack loads cleanly on kornia 0.8.3. Keep the pin only on an older pack.

Changed: SaveVideo quality

The CLI asks SaveVideo for crf 12 instead of its default 23, which smooths exactly the texture
these models add. The setting's location moved between ComfyUI releases, so it is read from the
server's own schema, and nothing is sent to a server that has none.

Measured: the shipped pixel conditioning against the int8 encoder

1.1 said the Q5_K_M tensors were "expected to be numerically close" to the int8 encoder's. They
are close, not identical: cosine similarity 0.934 on the tensor, and 48.3 dB PSNR between a
17-frame render using them and the same render through the int8 encoder. A cache is bit-identical
to the encoder that made it (PSNR inf, frame hashes equal, both modes). For exact parity with
your encoder, regenerate with --bootstrap-cond.

Also

• --setup checks the mode you will run (--model), including the fusion nodes and the right
  LoRA.
• Chunk sizing reserves the frames a chunk renders beyond its own (the 8n+1 rounding and refine's
  pad), so --budget is no longer overrun by up to a quarter; a chunk that reads past the end of
  the clip is padded; odd-sized sources are cropped one pixel to even dimensions in refine mode;
  and --budget 0 is refused instead of silently meaning "default".
• tools/check_release.py re-derives both size tables from the code, and asserts every refine
  fix in the shipped files.

4K demos attached below: forest river and band in a wheat field, refined to 3840x2144 in one pass each on an RTX 5090, with their sources. The comparison reel (33s, 1:1 pixels of the 4K frames) covers those two plus four more clips: a portrait clip refined in one pass, and three riding shots rendered in 2-second chunks, with each wipe kept inside one chunk.