Josia媒体保存

Every ComfyUI workflow ends in the same pile of output nodes: Save Image here, Save WEBM there, Save Audio for the soundtrack, VAE Decode before any of it because the sampler hands you a latent, not pixels. Josia媒体保存 collapses that pile into one node - and then does one thing the native nodes don't.

Why you'd bother

Three reasons, in order of how much they matter.

The latent gets written before the decode. Decoding is the memory spike; everything upstream is a tensor a fraction of the size. With 保存潜空间 on, this node saves the .latent first, then decodes. If the decode OOMs - what usually kills a big video or high-res run - the run isn't wasted: load it back with Josia加载Latent, free VRAM, decode separately.

You can skip the separate decode. Wire Latent plus a VAE in and it decodes internally, with 解码方式 (自动 / 直接解码 / 分块解码) covering the case where the plain decode won't fit. On a small card that's one node instead of two, and the tiled path replicates ComfyUI's VAEDecodeTiled semantics exactly (troubleshooting.md for why VRAM ceilings bite here). 图像 also passes straight through, so this drops in where Save Image already sits.

Inputs a beginner actually sets

filename_prefix defaults to JosiaMedia\Pic_%001% - writes to output/JosiaMedia/ as Pic_001, Pic_002. %001% is the counter placeholder, substituted in place: shot%0001%_raw gives shot0007_raw. %date%, %time%, %date:yyyyMMdd% and %time:hhmm% work too, same rules as ComfyUI's own prefix templating (image-io-metadata.md). Leave a counter out and you get an automatic _0001 suffix; same-name files never overwrite.

图像格式 lists 15 options, and formats your environment can't write are hidden rather than offered and then failing - AVIF, HEIF and JPEG XL only appear once a Pillow plugin provides them. A ⭐ before the name means that format carries workflow metadata. 视频容器 is 关 / MP4 / MKV / WebM / GIF / APNG / 动图WebP, synthesised from an image batch at the given fps; codecs the container can't hold (h264 into WebM) get swapped for its default instead of erroring. Audio is FLAC / WAV / MP3 / Opus with a bitrate ladder.

VAE1 is where this node plays nicely with its sibling: the default is 使用Josia模型加载VAE, which reuses whatever VAE the pack's model-loader node already loaded, with no wire. Wire the Video_VAE port and the dropdown greys out - wiring wins.

解码精度 (自动 / fp32 / fp16 / bf16) is the knob people never look for and always need: fp32 is the standard fix for black images and snow from a VAE decode (troubleshooting.md). 清理缓存 (关/轻度/深度, before or after decoding) reclaims only unreferenced tensors and CUDA free blocks - it never unloads your models, so repeat runs don't get slower.

Outputs: 图像 (your image, or the decoded result), 路径 (the main file's relative path, handy for a string node), Latent (pass-through, for two-sampler workflows), plus Video_Latent and Audio_Latent split out of a joint AV latent, None when there isn't one.

Install

ComfyUI Manager, search ComfyUI_JosiaNodes, or:

cd ComfyUI/custom_nodes
git clone https://github.com/Josia-doit/ComfyUI_JosiaNodes
# restart ComfyUI

Restart and check the console for a [JosiaNodes] ✅ line. There's no model to download - it uses your VAE. requirements.txt is empty, and that's mostly honest: image, audio and latent saving run on Pillow, torch and the standard library. Video is the exception - MP4/MKV/WebM go through PyAV, which ComfyUI's own video nodes install, so a full desktop install already has it. If you get 未安装 PyAV（av），无法写出视频, that's the missing piece, not a broken node.

The author's README is upfront that the pack ships mainly via GitHub and doesn't promise passage through the registry's automatic security scan, so if Manager can't find it, clone it manually. Normal custom-node rules: no review, no sandboxing (comfyui-ecosystem.md).

Common issues

"接了「Latent」但没有接「Video_VAE」，无法解码。" You fed it a latent with no VAE anywhere. Either wire one, or set VAE1 to 使用Josia模型加载VAE and make sure the model-loader node has actually run once.

Audio-only video models go half-silent. For LTX-2 / MiniMax-style av models the latent is a joint audio+video container, and both paths decode independently (ltx-video.md). No audio VAE set means the video saves and the audio is dropped - with an explicit console warning, not silently. Set VAE2, or wire Audio_VAE.

OOM on the decode. 解码方式 = 自动 already retries tiled after a CUDA OOM. To aim it, the tooltip's guidance is a 512 tile on 8GB, 256–384 on 6GB, with 64px overlap. For fragmentation, set 清理缓存 to 深度 with timing 解码前.

It re-saves on every queue. Correct behaviour - as an output node with a NaN change key, it must always run. The flip side: a quick iteration session quietly fills output/JosiaMedia/ with forty near-identical PNGs. Set 临时预览 on while you're fiddling; it writes a level-1 compressed copy to temp/, matching Preview Image.

One thing it won't do: the metadata it writes is ComfyUI's prompt/workflow block, not the A1111-style parameters string CivitAI wants (image-io-metadata.md).