RunningHub workflow page: https://www.runninghub.ai/zh-cn/post/2088976220955439106?inviteCode=rh-v1111
LTX 2.5 Fast Video Super-Resolution (Generative Detail)
Purpose: a video-to-video detailer/upscaler on the LTX 2.5 22B distilled transformer. Load a video, scale it up, VAE-encode it, run the LTX 2.5 detail pass with an IC-LoRA, decode with tiled VAE and save while keeping the original audio.
Graph (as saved)
20 nodes / 26 links / 3 groups.
- UNETLoader:
ltx-2.5-22b-distilled-transformer-bf16.safetensors; CLIPLoader:gemma4-12b-with-proj-ltx-2.5-comfy-int8-convrot.safetensors(typeltxv); VAELoader:ltx-2.5-video-vae-conv-bf16.safetensors. - LoRA:
ltx-2-19b-ic-lora-detailer.safetensorsat strength 1. Note inconsistency: the graph's Markdown note names a different IC-LoRA (ltx-2.3-22b-ic-lora-instant-shave-0.9.safetensors); the loader node is the saved state and is what this listing quotes. - Sampler:
LTXVLoopingSamplersaved with its own widget set (56 / 24 / 1 / 0.5 / 1 / 1 / 1 / 1 / 0 / 0 / 1000 / "0"),KSamplerSelecteuler,ManualSigmas,CFGGuider,LTXVConditioningframe rate 24. - I/O: LoadVideo → scale (RTX upscale node; the node title says ×1.5 while the saved multiplier widget is 2.0 — check it before queueing) →
VAEEncodeTiled→ decode viaLTXVSpatioTemporalTiledVAEDecode(4/4/16/4 tiles) → CreateVideo 30 fps → SaveVideo prefixLTX-2.5/V2V_Detailer_Upscale.
How to use
Load the source clip, set the scale factor (the note says the default is 1.5× and that scale is the only thing you normally change), and run. If you change the multiplier, re-check VRAM and the tiled decode settings.
Limits and not-verified items
No inference was run for this listing: no speed, VRAM, fidelity or temporal-consistency claim is made. The LTX model, IC-LoRA and VAE are not bundled; check their licenses separately. The graph notes carry third-party attribution (MarkDKBerry.com YouTube/Patreon) — preserved as-is, and this listing does not claim to be the original source. The cover image is a promotional still, not a frame produced by this graph.
Tutorial: YouTube · Bilibili. Workflows supplied through AI-KSK; verify workflow, node, model, LoRA and cover rights separately. No blanket commercial or redistribution license is asserted.
Description
RunningHub workflow page: https://www.runninghub.ai/zh-cn/post/2088976220955439106?inviteCode=rh-v1111
LTX 2.5 Fast Video Super-Resolution (Generative Detail)Purpose: a video-to-video detailer/upscaler on the LTX 2.5 22B distilled transformer. Load a video, scale it up, VAE-encode it, run the LTX 2.5 detail pass with an IC-LoRA, decode with tiled VAE and save while keeping the original audio.
Graph (as saved)20 nodes / 26 links / 3 groups.
- UNETLoader:
ltx-2.5-22b-distilled-transformer-bf16.safetensors; CLIPLoader:gemma4-12b-with-proj-ltx-2.5-comfy-int8-convrot.safetensors(typeltxv); VAELoader:ltx-2.5-video-vae-conv-bf16.safetensors. - LoRA:
ltx-2-19b-ic-lora-detailer.safetensorsat strength 1. Note inconsistency: the graph's Markdown note names a different IC-LoRA (ltx-2.3-22b-ic-lora-instant-shave-0.9.safetensors); the loader node is the saved state and is what this listing quotes. - Sampler:
LTXVLoopingSamplersaved with its own widget set (56 / 24 / 1 / 0.5 / 1 / 1 / 1 / 1 / 0 / 0 / 1000 / "0"),KSamplerSelecteuler,ManualSigmas,CFGGuider,LTXVConditioningframe rate 24. - I/O: LoadVideo → scale (RTX upscale node; the node title says ×1.5 while the saved multiplier widget is 2.0 — check it before queueing) →
VAEEncodeTiled→ decode viaLTXVSpatioTemporalTiledVAEDecode(4/4/16/4 tiles) → CreateVideo 30 fps → SaveVideo prefixLTX-2.5/V2V_Detailer_Upscale.
Load the source clip, set the scale factor (the note says the default is 1.5× and that scale is the only thing you normally change), and run. If you change the multiplier, re-check VRAM and the tiled decode settings.
Limits and not-verified itemsNo inference was run for this listing: no speed, VRAM, fidelity or temporal-consistency claim is made. The LTX model, IC-LoRA and VAE are not bundled; check their licenses separately. The graph notes carry third-party attribution (MarkDKBerry.com YouTube/Patreon) — preserved as-is, and this listing does not claim to be the original source. The cover image is a promotional still, not a frame produced by this graph.
Tutorial: YouTube · Bilibili. Workflows supplied through AI-KSK; verify workflow, node, model, LoRA and cover rights separately. No blanket commercial or redistribution license is asserted.