In the previous 18 articles, I told you about the machines, the Ventura, the TowerNet, LoRA, and a list of downgrades for you. As soon as the article came out, the first person in the comment section said:
"It's not the memory that creates the explosion. What's the use of your list? I'm sorry
I had a moment because he was right. It's my first time ComfyUI Runchart video, a five-second little piece, ran to half the whole software card. Cut to the task manager32G, MEMORY FULL. THE SYSTEM KILLED THE PROCESSI don't know. At first I thought it was a graphic cardIT'S RAM, NOT VRAM。
So this one is about the memory of video production. And just in time, ComfyUI is v0.29 It quietly changed one place and filled most of the pit。
I. Video decoding. Why eat memory like this
YOU GENERATE A VIDEO, AND THE LAST STEP IS: VAE DECODES THE SUBSPACE INTO ONE FRAME, THEN THESE FRAMESencoding as mp4 into hard drive.
The old version is stupid: it starts with theEvery frameALL ARE STORED IN THE SYSTEM (RAM), AND ONCE ALL THE FRAMES ARE IN PLACE, ANOTHER ENCODED DISK. THE PROBLEM IS ON THIS "CORRESPONDING" -- THE MINUTE THE MEMORY IS BLOWN UP AS SOON AS THE FRAME NUMBERS X RESOLUTION GO UP。
| take | Memory pressure (estimate, unrun) | clarification |
| 5 SECONDS . 480P | Small | Less frame, less pressure |
| 10 SEC 720P | Significant increase | Frame x resolution takes off, memory goes up |
| Video input (Turson video reference) | Same memory | The decoded reference video will also buffer the entire frame |
And it's not just the "out of video" step. You use itVideo inputThe decoded reference video also buffers all frames into the memory, with the same size。
ii. v0.29 changed: video flow code
ComfyUI v. 0.29.02026-07-29 published) In the chapter of Performance & Stability, this section has been specially adapted. Official PR describes it as:
Video transport (PR #14813)
It's just a translation:The video transcoder was converted to a fluid treatment - each frame is no longer buffered into memory。It's called "Salvaging All Frames and Writing Discs", it's now "Generating One Frame, Writing One Frame, Release One Frame" and the memory always leaves only a small amount of frame currently being processed. Long video, high resolution benefits the most, as memory occupancy no longer surges with frames。
along with v0.29, there are two other things related to memory/ resource that deserve to be mentioned:
1. BACKSTAGE MEMORY LIMIT DEFAULT REFERS TO 128GB(PR #15068) The upper limit of the CofyUI backstage occupancy system is raised, and large memory machines are less likely to be crushed by it。
2. Comfy Kitchen More Int8 Optimise(PR #14980): Int8 quantifying this road earlier (v 0.27 enters the main line), v0.29 is continuing to grindIt's not new。 。 。 。 。 。 。 。 。 。 。 。 。 。 。 。 。 。 。 。 。 。 。 。 。 。 。
Integrity:IT'S CALLED "VIDEO FLOW-CLASSED MEMORY" WHICH I SAW AS A MECHANISM IMPROVEMENT FROM THE OFFICIAL PR DESCRIPTIONWe're not running on the cardSuper-long video was compared back and forth, with no public figures for the specific drop. But in principle, long video, high-resolution scenes benefit most; you can feel it when you upgrade。
iii. Easy to mix: This is saving memory, not saving memory
Part 18 What about ComfyUI, from installation to mastery? A list of downgrades by priority IT'S VRAM, IT'S MEMORY. THEY'RE NOT THE SAME:
| Symptom | Gene | Which one |
| CUDA out of memory | MODEL / VAE CANNOT LOAD VISIBLE | List of downgrades |
| Process systematically killed / card killed | Frame buffer full memory | title (iv. 0.29.2) |
Video generation is a double-pressure "Standing + Memory" scene. A lot of people just show up, and they fall, and it's the half of the memory。
clean loop: upgrade to v.29.2 and run a long video (do it now)
Let's get to the point. This one's real. After you've got another video, "Doesn't fail under a long video," you'll be able to confirm that the current code actually works。
Preconditions: You've been running videos to generate workflows (Van, HunyuanVideo, or trinking nodes). Dropped: Basic Textile Chain (Page 05) + Video-related Node (Load Video / VAE Encode / Decode + Video Encoding Node)。
Target: Raise ComfyUI to v.29.2 and rerun a video of "intentionally lengthening" with a more stable memory occupancy and no further collapse。
Step (sight, don't jump):
1. Update ComfyUI: Desktop Dot "Check Update" upgrade; Git release restarts after ComfyUI directory guit pull。
2. Confirm Version: Menu Help → About, see version number V0.29.2 (2026-07-31 latest stable version)。
3. Rerun Video Workstream, THIS TIME INTENTIONALLY PULLS OUT MEMORY PRESSURE FOR A LONG TIME OR AT A RESOLUTION (5 SECONDS →10 SECONDS, OR 480 P →720P)。
4. Look at the memory curve: With the task manager on, the old version will increase steeply at the end of the generation, and the new version should be much more flat and no longer flush。
Checkpoint
1 About displays v.29.2; long video runs and dies
2 Common error = an updated plugin with an error "Node Not Found" plugin needs to be reloaded / updated (updated in Manager); or an additional 18 downgrade list is not visible enough
3 still collapses, either memory or display double pressure, to see part 18 in the half; memory still closes to start-up parameters –cache-none
4 Video came out but the resolution/ step question is not a transcode。
Job (chosen): The same video, one before and after the upgrade, tells me what the difference is in the memory peak -- I'm curious to know how much you can save。
iv. Parameter speed sheets (video memory relevant, copying)
| take | Recommendations | Remark |
| SHORT CLIP (<5 SECONDS 480P) | It doesn't make any difference | Less frame, less pressure |
| Long Snippet (>10 seconds) | i have to go to v.29.2 | Fluid transcode yields the most |
| HIGH RESOLUTION (720 P+) | ig v0.29.2 + 18 visible decline | Memory + Visible Double Pressure |
| Video input (Turson video reference) | l v.29.2 | It's also good to decode |
| Memory remains tight (extreme environment) | add -cache-none | Uncut intermediate results, save memory exchange speed |
v. Contrast of start-up parameters (RAM, copying)
| parameter | corresponds English -ity, -ism, -ization | For what |
| (default) | v0.29 streaming code is open | up to v. 0.29.2 |
| –cache-none | No Cache Node Results | The memory is on the bottom |
| –reserve-vram 1.5 | Save System Visibility | When double pressure is visible / memory (see under 18) |
One sentence:only upgrades to v. 0.29.2, transcode automatic flow; until memory limits add –cache-none。
VI. 4060 / 8G SPECIALIZED & I STEPPED ON THE PIT
4060 THIS CARD, VISIBLE 8G, MEMORY USUALLY 16~32G. VIDEO GENERATION IS THE DOUBLE PRESSURE OF THE "DIGNO REPOSITORY MODEL + MEMORY CACHE FRAME": TO SEE PART 18 FOR THE HALF, THE HALF FOR THE MEMORY。
Trail experience:I ran a 10-second graphic video, only adjusted the visual parameters, the memory was undone, and the frame buffer was 32G full, the system killed the process. Then I went up v. 0.29, and the resolution dropped from 1080 P to 720 P, and the same thing went through. So..Video blast memory, first version, then resolutionDon't die on display。