Video upscale workflow for Low VRAM (Extreme F… v4.1

The Goal: Help low VRAM users (under 12GB) push their videos to 1080p and beyond—keeping the motion smooth, the quality sharp, and the audio perfectly synced without breaking your GPU.

V4.1 Minor Update

Use Model for Ratio Upscaling instead of RTX.

The final resolution will be a little larger than 1080p.

Fine tune the workflow process to SeedVR2/VOSR2 => Upscale w/ Model => RTX upscaler.

Update Quick guide and download link.

Non-Low VRAM Users: If your hardware environment is ideal, you can directly Skip (Bypass) RTX and upscale to your target multiplier directly during the previous SeedVR2 / VOSR2 phase.
💡 Resolution and Aspect Ratio Adjustment Tips
* Resolution Reference: Taking Minimax H3 as an example, the native video is 480 x 864 (9:16). After the secondary upscaling process mentioned above, it reaches 1152 x 2072, which is slightly larger than standard 1080P.
* Flexible Configurations: If you have an ideal hardware environment (non-Low VRAM users), you can directly upscale to your target factor during the SeedVR2 / VOSR2 stage.
In this scenario, Upscale with model is used purely for denoising or repair, and RTX Upscale can be completely turned off.
Furthermore, you can choose not to reduce the frame rate during the Load Video stage, upscaling directly at the original 24 FPS (simply change the force_rate parameter from 12 to 24 in the Load Video node).
* Non-Standard Aspect Ratios: If the original video is not in a 9:16 aspect ratio, remember to manually adjust the ratio_preset appropriately.
非低 VRAM 用戶：若硬件條件理想，可直接跳過RTX放大，並在先前的 SeedVR2 / VOSR2 階段直接放大至目標倍數。
💡 解析度與比例調整技巧
* 解析度對照：以 Minimax H3 為例，原生影片為 480 x 864 (9:16)，經過上述二次放大後會來到 1152 x 2072，略大於標準 1080P。
* 靈活搭配：若環境理想（非低 VRAM 用戶），可以直接在 SeedVR2 / VOSR2 步驟放大至目標倍數。在此情境下Upscale with model 僅需用作去噪或修復，而 RTX Upscale 則可以直接關閉。甚至在 Load Video 階段也無需降低幀率，可直接以原片的 24 FPS 進行放大（只需在 Load Video 節點中將 force_rate 由 12 修改為 24 即可）。
* 非主流比例：若原影片並非 9:16 比例，請記得手動適當調整 ratio_preset。

V4.0 - Major Update

Add SeedVR2 to Upscale option.

Move Frame Interpolation group to independent subgraph.

decrease the force rate to 12FPS in Load Video.

Update quick guide and download link.

Add Model selection Node (Green Node)

1 = VOSR2

2 = SeedVR2

Drop Frame Rate to 12 FPS During "Load Video"

Adjust the force rate parameter in the *Load Video** node to 12.

* Strategy: To prevent OOM (Out of Memory) errors and shorten upscale processing times, the frame rate is cut in half to 12 FPS upfront before enhancement and upscaling. During the final frame interpolation phase, it will be quadrupled back up to 48 frames.

For SeedVR2, try to use the INT8 convrot format paired with Comfy Kitchen attention. This generates twice as fast as the GGUF version. (You can also choose Sage attention, but make sure to set the Backend attention to pytorch attention).

Upscale with Model could be switch OFF here if you don't need it.

V3.0 - Major update to the workflow

Add VOSR2 to to enhance the quality.

Use RTX to upscale the video to 1080p.

Frame interpolation to mulitply the frame rate.

Add Switch button on the Subgraph panel for each enhance step.

Update quick guide in the workflow.

Make sure switch OFF anything unwanted before click RUN.

VOSR2 is a brand-new tech for boosting image quality. While made for photos, it works like magic on video. That said, trying a massive upscale right out of the box will almost 100% crash your GPU with an OOM (Out of Memory) error if you're under 12GB VRAM. (A 15-second 1080p clip could literally take over an hour).

【The Hack】: First, shrink your 480p video by 50%, then let VOSR2 scale it back up by 2x.

* The Result: You keep the exact same resolution but get a massive jump in clarity and details.

* Render Time: Around 4 minutes total—totally worth it.

* 💡 Pro-Tip for Saving VRAM (Tile Settings):

The *Tile sizing** in this node (defaults to 512 / 64 / 512 / 64) chops up the frame during generation to save your graphics memory.

* If your card is still struggling, you can drop these values down to 256 / 32 / 256 / 32. Just keep in mind that this will hit your final image quality.

Honestly, *sticking to the default 512 gives the absolute best results** based on testing.

VOSR2 是目前用來提升畫質的新技術（原用於相片，拿來修復影片效果也很棒）。但對於低 VRAM 用家來說，直接用它來做「高倍率放大」極不現實（12GB VRAM 以下幾乎 100% 會 OOM 記憶體崩潰，生成 15 秒 1080p 影片可能耗時超過 1 小時）。

【折衷方案】：先將 480P 影片**縮小 50%**，再用 VOSR2 放大 2 倍。

* 效果：維持原解析度不變，但畫質得到顯著修復與提升。

* 耗時：總生成時間約 4 分鐘，在可接受範圍內。

* 💡 Low VRAM 關鍵節點參數設定：

節點中的 *Tile 參數**（預設值為 512 / 64 / 512 / 64）是用於生成時將畫面切塊處理以節省 VRAM。

* 如果依然遇到 VRAM 不足，可將其調低至 256 / 32 / 256 / 32，但此舉會犧牲部分畫質。

根據目前測試，*維持預設的 512 效果最為理想**。

SWITCH IT OFF on the panel if you feel too slow in generation, and everything unchange.

如果生成速度太慢, 在面板關掉它, 然後用回先前的方案。

Special thanks to the node creator: ylchen333👏

V2.1 - Minor update to the workflow

add switch to all three function among RTX/ upscale with model/ frame interpolation

add reserved VRAM to avoid OOM

For Low VRAM user, if you meet OOM in upscale 15s video, you might disable upscale with model and scale x 2.25 in RTX, megapixels = 3.7 and continue the frame interpolation.

i.e. your original video resolution x 2.25 = almost 1080p

most low VRAM user will use 480p(0.4mp) as the orginal resolution during generation

and 1080/480 = 2.25

of course this figure you might fine tune by yourself.

V2.1 - 工作流小幅更新

為 RTX / 模型放大（Upscale with model） / 補幀（Frame interpolation）這三項功能全部加入獨立開關。
增加預留顯存（VRAM）功能，以避免記憶體溢出（OOM）崩潰。
針對低顯存（Low VRAM）用戶的建議：
- 如果在放大 15 秒的影片時遇到 OOM 崩潰，可以嘗試關閉「模型放大」。
- 並在 RTX 中將縮放倍率設為 2.25 倍（百萬像素 Megapixels = 3.7），然後繼續進行補幀。
- 註：您的原始影片解析度乘以 2.25 倍，大約就等於 1080p。
- 大多數低顯存用戶在生成時，會使用 480p（0.4mp）作為原始解析度。
- 而 1080 除以 480 剛好等於 2.25。
- 當然，這個數值您也可以根據自己的情況進行微調。

V2.0 - Major update to the workflow

add frame interpolation function in the subgraph

exchange the upscale order.

add several button on the subgraph interface let user make their option.

frame interpolation could be switch off by this button as shown below.

provide link to download interpolation model.

V2.0更新

加入frame interpolation 功能

更改放大的順序由先完成RTX改成完成upscale with model 後再做RTX修復

在subgraph中增加frame interpolation開關提供用家選擇

frame interpolation可以用下方圖中的按鈕開關

提供下載連接

V1.1 - Minor update to the workflow.

Minor update to the workflow.

Add last frame generate after video upscaled.

This was supposed to be included in the workflow for the next update of Minimax H3 workflow, but I feel like some users would just upscale every single output right after generating it.

小更新

在影片upscale後加入自動生成last frame.

本來應該加在Minimax H3下次更新的Workflow中

但總覺得有用家會每生成一條就立即upscale

Simple workflow for Low VRAM user. (Extreme FAST)

This workflow isn't limited to MiniMax H3 users—those running Wan2.2 and LTX-Video 2.3 can use it exactly the same way.

Just a quick heads-up: if your video has no audio, remember to disconnect the Audio link from the Load Video node, otherwise you'll get an error.

Once again, what I'm sharing today is an ultra-simple workflow. It follows my usual style: simple, practical, completely spaghetti-free, and 100% Low VRAM friendly.

I’ve covered the rest of the details directly inside the workflow, so I won't repeat them here.

Low VRAM專用放大影片Workflow. (極快)

不限於Minimax H3 用家, Wan 2.2 和LTX 2.3的用家同樣可用

沒聲音的影片記得斷開Load Video 中Audio的連接, 不然會報錯

今次分享的亦是超簡單Workflow。

亦遵循我一貫作風, 簡單, 實際, 不會意粉化, Low VRAM能用

其他詳細在workflow中有說明就不再重覆了

How to merge videos with FFmpeg in ComfyUI Desktop (2026)?

https://civarchive.com/articles/34119/how-to-merge-videos-with-ffmpeg-in-comfyui-desktop-2026