Video Upscale

Video Upscaling & Restoration

Overview

"Upscaling video" in ComfyUI splits into three jobs, and the quality win comes from doing them in the right order:

1. Spatial restore + upscale. A temporal-aware model that increases resolution AND cleans compression blocks, blur, and AI-gen mush while keeping frames consistent over time. This is the part a plain image upscaler (ESRGAN, UltimateSDUpscale per-frame) does badly. Per-frame upscalers flicker because each frame is sharpened independently. Use a video model.
2. Frame interpolation (VFI). Synthesize in-between frames to raise fps (e.g. 24→48/60) for smooth motion. Do this after the spatial pass.
3. Encode. Mux frames (plus original audio) back to an MP4.

The two leading temporal restorers in 2026 are SeedVR2 (diffusion-transformer restorer, the proven workhorse) and FlashVSR (newer one-step streaming VSR, faster). Frame interpolation is RIFE (or FILM) via ComfyUI-Frame-Interpolation.

Verification note: every node/pack/model name below was confirmed against the GitHub repos and the ComfyUI registry / Manager as of June 2026. Where a name is approximate or version-dependent it is flagged. Do not substitute a node you can't confirm is installed. Check with install_custom_node (action: "list") / create_workflow (action:"node_info").

⭐ Recommended current pipeline (2026)

Downscale → SeedVR2 (temporal restore+upscale) → RIFE (interpolate) → VHS encode. This modernizes the user's classic recipe (below) with the current SeedVR2 node pack and is the path to ship by default. FlashVSR is the faster alternative for the restore stage (see "FlashVSR" section). Swap it in when speed matters more than absolute fidelity.

Node-graph sketch

LoadVideo  ─► GetVideoComponents ─► (IMAGE frames, audio, fps)
                       │
                       ▼
        ImageScaleBy / ImageScale   ◄── DOWNSCALE first (e.g. 0.5×) — clean,
                       │                  small input for the restorer
                       ▼
        SeedVR2 Video Upscaler  ◄── DiT model + VAE + (block swap) + (tiling)
           ├─ "SeedVR2 (Down)Load DiT Model"
           ├─ "SeedVR2 (Down)Load VAE Model"
           └─ ["SeedVR2 Torch Compile Settings"]  (optional speedup)
                       │  (restored, high-res frames)
                       ▼
        RIFE VFI (4.0 - 4.9)   ◄── multiplier 2 (e.g. 24→48 fps)
                       │
                       ▼
        CreateVideo (fps = source × multiplier) ─► SaveVideo
        — or — VHS_VideoCombine (carries audio passthrough)

LoadVideo / GetVideoComponents / CreateVideo / SaveVideo are core ComfyUI video nodes (same ones the official comfy.org SeedVR2 template uses). VHS_LoadVideo / VHS_VideoCombine come from ComfyUI-VideoHelperSuite (installed) and are preferred for the final encode because they pass the original audio through.

Why downscale FIRST (the load-bearing trick)

• The restorer wants a clean low-res input, not a big dirty one. SeedVR2 (and FlashVSR) regenerate detail. Feeding them a small frame forces the model to synthesize sharp detail rather than faithfully magnifying existing compression artifacts and noise. Downscaling first averages away block noise, so the restorer hallucinates clean, coherent texture.
• VRAM and speed headroom. Cost scales with input pixels × frames. Halving each dimension is ~4× fewer pixels per frame, which buys you a larger temporal batch (the thing that kills flicker, see below) and a bigger target multiple.
• It turns "upscale" into "restore-and-upscale." A 720p source downscaled to 360p then SeedVR2'd to 1080p+ looks much better than 720p→1080p straight, because the model rebuilds rather than stretches.
• Rule of thumb: downscale to 0.5× (or to a ~360 to 480p short side) for messy, low-bitrate, or AI-gen footage; skip the downscale for already-clean, high-bitrate sources where you only want more pixels.

Quick local path (no downloads) — per-frame ESRGAN + built-in RIFE

When the user wants a fast result on what's already installed (no SeedVR2 / FlashVSR multi-GB download), use the ESRGAN upscale models most setups already have. Check first with list_local_models (common ones: 4x_foolhardy_Remacri, the best for realistic footage/water/skin, and 4x-ClearRealityV1 for clean/sharp).

• Upscale: ImageUpscaleWithModel with a 4× ESRGAN model, then ImageScale back down to a clean 2× (a 4×→2× supersample). That downscale-after step is the single biggest quality lever here. It averages out per-frame noise.
• Interpolate: the built-in FrameInterpolate with RIFE v4.26 (see the Frame-interpolation section; no custom node on 0.26+).

Tradeoff: flicker. ESRGAN upscalers are per-frame (no temporal awareness), so they can shimmer or flicker on video, most visible on water and fine detail. The 4×→2× supersample mitigates it; if it still shimmers, that's the signal to switch the upscale stage to a temporal model (SeedVR2 / FlashVSR below), which is the real fix. So: per-frame ESRGAN = quick and local; SeedVR2 = flicker-free and best. Order is unchanged: upscale the real frames first, then interpolate.

This is the right default for a "do it now, locally" request; reach for the temporal restorers below when quality (or zero flicker) matters more than turnaround.

SeedVR2 (recommended restore stage)

Pack: ComfyUI-SeedVR2_VideoUpscaler (author numz), GitHub numz/ComfyUI-SeedVR2_VideoUpscaler, installable via ComfyUI-Manager / registry by that name. Install with panel_install_node or apply_manifest.

Node classes (4):

Models (auto-download to models/SEEDVR2/; 3B = lighter, 7B = best quality):

Key params on "SeedVR2 Video Upscaler":

3B vs 7B: start with 3B fp8. It's the speed/quality sweet spot for most footage. Move to 7B only when you need maximum reconstruction on faces/text and have the VRAM (or use the 7B GGUF + block swap).

FlashVSR (newer SOTA — faster restore stage)

Pack: ComfyUI-FlashVSR (author 1038lab), GitHub 1038lab/ComfyUI-FlashVSR, built on FlashVSR V1.1 (one-step diffusion + locality-constrained sparse attention + tiny conditional decoder). Manager name ComfyUI-FlashVSR. Models auto-download from HF 1038lab/FlashVSR to models/FlashVSR/ on first run.

Note: several community forks exist (smthemex/ComfyUI_FlashVSR, lihaoyun6/ComfyUI-FlashVSR_Ultra_Fast, naxci1/ComfyUI-FlashVSR_Stable). The 1038lab pack is the cleanest two-node implementation; pick a fork only if you need its specific VRAM tricks.

Node classes: FlashVSR ⚡ (preset: Fast / Balanced / High Quality) and FlashVSR Advanced ⚡ (model_version = Tiny / Tiny Long / Full, enable_tiling, speed_optimization, quality_boost, sageattention). Supports 2x and 4x (4x recommended). Needs ≥21 input frames. Variants: Full (best, heavy VRAM), Tiny (fast), Tiny Long (low VRAM). SageAttention adds ~20 to 30% speed.

When to prefer FlashVSR over SeedVR2: real-time, long clips, or speed-critical jobs, or when SeedVR2 is too slow on your hardware. Prefer SeedVR2 when you want the strongest restoration of badly-degraded footage and can spend the time.

Frame interpolation (VFI)

Built-in (ComfyUI 0.26+) — PREFER this, no custom node needed

Current ComfyUI ships a core frame interpolator, nodes FrameInterpolationModelLoader + FrameInterpolate (RIFE/FILM), so on 0.26+ you do NOT install a custom node. It auto-detects checkpoints dropped in models/frame_interpolation/ (that folder is empty by default, which is why the model dropdown looks blank). Core-compatible weights live at HF Comfy-Org/frame_interpolation (under a frame_interpolation/ subpath):

Drop a file into models/frame_interpolation/ and restart so the dropdown populates (ComfyUI caches model lists). download_model may not target that folder; pull it directly into models/frame_interpolation/. Check what's there with list_local_models / the core node's dropdown before installing anything.