I could not use the old Supir as of May 2026. In June got help from Kijai with a new workflow that works. I rather assumed that someone would upload it here, but since nobody has, I shall.

Supir is, in my experience, by far the best upgrade model for images.

Installation:

Base model: juggernautxl_v9rdphoto2lightning

Found here: juggernautXL_v9Rdphoto2Lightning.safetensors · AiWise/Juggernaut-XL-V9-GE-RDPhoto2-Lightning_4S at main

Put it in models\checkpoints\

There are two SUPIR models to use with this. One is v0Q_fp16, where Q stands for Quality, giving it more flexibility to invent whatever it thinks is missing. The other is v0F_fp16, which adheres more strictly to the original image.

Found here:

Kijai/SUPIR_pruned at main

You put those two in

\ComfyUI\models\model_patches

If I got anything right, kudos to Kijai.

Whatever I got wrong, please correct me in the comments below.

Configuration of the two nodes:

FlashVSR:

For almost everything I use mode= tiny , scale=2

but for the highest quality, consider this:

🟦 FlashVSR Mode Differences (Tiny vs Full vs Tiny‑Long)

1. Tiny

Purpose: Maximum speed Quality: Lowest of the three VRAM: Lowest Use case: Real‑time, previews, fast pipelines

Technical behavior:

• Uses the tiny conditional decoder (lighter reconstruction path)

• Reduced attention range → faster but less temporal stability

• Ideal for streaming or quick batch processing

• Recommended when processing time matters more than detail

2. Full

Purpose: Maximum quality Quality: Highest VRAM: Highest Use case: Final renders, best temporal consistency, best reconstruction

Technical behavior:

• Uses the full temporal attention and full decoder stack

• Best edge recovery, best motion handling

• Highest fidelity reconstruction (especially with FlashVSR v1.1)

• Recommended when speed is not critical and GPU VRAM is sufficient

3. Tiny‑Long

Purpose: Long videos + low VRAM Quality: Slightly better stability than Tiny VRAM: Very low Use case: Long sequences, low‑VRAM GPUs, stable lightweight inference

Technical behavior:

• A special variant introduced in FlashVSR v1.1 for efficient long‑sequence inference

• Reorganized weights + optimized decoder for long videos

• Uses:

  • TCDecoder.ckpt (temporal consistency decoder)

  • posi_prompt.pth (KV‑cache stability)

  • Wan2.1 VAE modules for efficient encoding/decoding

• Designed for deployment, not training

• In ComfyUI, explicitly described as: “Tiny Long (Low VRAM): optimized for GPUs with limited VRAM, ideal for long videos.”

Tiled

tiled_vae= reduces VRAM, lowers speed

tiled_dit= reduces VRAM, lowers speed

Video Combine:

For almost everything I use 10MB/s and AV1,

but for the highest quality, consider this:

Recommended bitrates (H.264 / H.265 / AV1)

1080p

Low: 8–12 Mb/s

Good: 15–20 Mb/s

High: 25–35 Mb/s

1440p

Good: 20–30 Mb/s

High: 35–50 Mb/s

4K

Good: 35–50 Mb/s

High: 60–80 Mb/s

Archival: 100+ Mb/s

AV1 is more efficient, but 10 Mb/s is still too low for 1080p+.

The best quality is Prores / 4444

but the files run into Gigabytes instead

of Megabytes, and CIVITAI will not take it.

You can always crunch it down to AV1 with shutter encoder.

yuv420p or p010le?

yuv420p

8‑bit

Universal compatibility

Lowest quality (banding, less gradient detail)

Good for web delivery, old players, H.264 baseline

p010le

10‑bit

Semi‑planar (NV12‑style)

Preferred by NVENC, QuickSync, AMD VCN, Vulkan

Best for AI video, upscaling, HEVC/H.265, AV1

Much smoother gradients, fewer artifacts