There's 17 in front ComfyUIIt's not like you're in the middle of something like this. But there's a Can, and almost every new guy hits, and I don't go around
When I first tried to run FLUX, I got excited to drag into the workstream, a little Queue Prompt, a key to the interface, and the console painted a big red CUDA out of memoryI don't know. I thought the card was broken, reloading, quenching, all night. I finally found out that it wasn't brokenIt doesn't even fit.
And then I realized that the ComfyUI newspaper OOM, not "you can't computer," was "the peak of this moment's presence." And this peak, yesThey can be removed, downgraded, distributedYeah. It gives you a list of downgrades in order of priority — from the most effective to the most extreme, and then down, basically running。
Remember:OOM IS NOT AN INCURABLE DISEASE, IT'S A "PEAKS TOO HIGH" AND CAN BE BROKEN DOWN。
I. WHERE'S THE STOCK
A map is going from noise to product, and several things are going to be stuffed in it. The biggest, the first. In order of size, you've got it in your mind:
| Quite a mess | Approximate scale (estimate, unrun) | clarification |
| Model weight (FLUX fp16) | ~23 GB | biggest head, fp16. fullest accuracy |
| T5 TEXT ENCODER (FLUX) | ~9 GB | Invisible! A lot of people just forget it |
| resolution (20482) | ~8 GB | Resolution square, double it |
| VAE DECODING (20482) | ~8 GB | High-intensity map decoders are full at once |
| watch / ControlNet / Detailer | EACH ~2-3 GB | Add a branch and a branch |
| Cache / Preview | ~0.5-1 GB | Small head, but small |
Trail experience:My first drop, I replaced the bottom model with a fp8, and it exploded. Half a day of finding -- T5 or fp16, nine G squirms. Change the T5 to a full run. Don't step on this pit。
So the core line of the downgrading is:START WITH THE BIG HEAD. DON'T PUSH ON THAT CACHE OF 0.5 G。
II. LIST OF DOWNGRADINGS BY PRIORITY (P0 → P3)
HERE'S THE WHOLE SOUL. P0 IS THE QUICKEST AND THE FIRST TO TRY; P3 IS THE LAST TO MOVE。
| Priority | Peak source | Degrade Actions | Command / Node |
| P0 | Model weight (maximum head) | Fp8 / GGF Quantification | Load Diffusion Model node weight_dtype=fp8_e4m3f; or .gguf With UNnet Loader (GGF) |
| P0 | T5 ENCODER (FLUX) | Change fp8 / GGF Version T5 | DualCLIPLoader (GGF) Load Quantification T5 |
| P1 | lentent resolution | Down to 10242 or 5122 | Emtty Latent |
| P1 | VAE CODE | Block decode | VAE Decode (Tiled), tile_size=512, overlap=64 |
| P2 | watch / ControlNet | closed branch, bs = 1 | Workflow nodes |
| P2 | Video frame x resolution | Frame Down + Time Segment | Temporal_size / Temporal_overlap |
| P3 | Cache / Preview | Off Preview, Off Cache | –preview-method none, –cache-none |
It's simple:Change the model format (P0), then lower the resolution, Tiled (P1) and finally move the cache (P3)。The reverse is half the work。
iii. a p well to be clear: –lowvram may be added to the new edition
You search the old school, there's eight in ten –lowvram“. But I'm gonna pour cold water..In the current ComfyUI, this is probably no use.
i compared the official start-up parameter document (v.2/29.2) to the official start-up parameter document–lowvram The text reads as follows:
No effect when dynamic VRAM is available.
Translate it:Dynamic VRAM is not effective when it's on。The current stable version of ComfyUI on the Nvidia graphic cardDynamic VRAM default oni don't know. in other words, the –lowvram that you're so excited about is an empty operation in the new edition。
Dynamic VRAM is the new version of the ComfyUI "Auto-Creative Manager": Models move as much as they need to be returned to the system, much smarter than the old "lowvram Force Part". So for most 4060 usersDon't fight with –lowvram. Let Dynamic VRAM take care of it. Just drop the model format and resolution(See sections II and V)。
Add:You want to force the old-fashioned –lowvram? But I don't usually suggest that--Dynamic VRAM is much more economical. On the other hand, –normalvram, –medvram: I looked at the official start-up parameter page, and I thought that it would be a good idea for me to be able to do that- medvram doesn't exist–normalvram is not listed in the new version (old documents are available, versions are different), both untrusted and not added。
version dividend: upgrade to 0.29.x optimization of white pick-up
if you're still in the old version, it's a push for your promotion -- the new version has two solids of low-end machines:
1. Dynamic VRAM AutoManaging(The current stable version is defaulted on Nvidia). As I said earlier, it's a modern version of the C.O.C. Manager, smarter than manual-lowvram. Upgrade is good, not good。
2. Video Streaming Code(v0.29.0 introduced). Official PR describes the video switch code from "Cache memory of every frame" to "stream processing" and the memory pressure has dropped significantly. This is critical for people who do video work and have insufficient memory。
Description:THE SECOND POINT I SAW FROM THE OFFICIAL PR DESCRIPTIONWe're not running on the cardYou can feel it when you upgrade。
Clean-up ring: 10 minutes, your 4060 with a low-visibility scheme
LET'S GET TO THE POINT. THIS ONE'S REAL. AFTER YOU HAVE AN EXTRA FLUX RUN THROUGH THE 8G DISPLAY, YOU CAN CONFIRM THAT THE DOWNGRADING ACTUALLY WORKED。
Preconditions: You've got the basic graphic workflows of Book 05 (Load Checkpoint / Load Diffusion Model → Two CLIP Text Encode → KSampler → VAE Decode → Save Image). If you lose it, press this six nodes。
Target: Replace a FLUX fp16 figure that was originally on OOM on 8G with a fp8 figure, which is visible and visible。
Step (sight, don't jump):
1. Change Model Format_Other Organiser Select fp8 version - Set the weight_dtype in node as fp8_e4m3f; or directly select a .gguf quantitative model。
2. DON'T FORGET, T5If FLUX is used, CLIP/T5 load is also replaced with a quantitative version (DualCLIPLoader (GGUUF) load T5). It's invisible。
3. Decrease Resolution: Empty Latet's width dropped from 15362 or 10242 to 5122 or 7682 first, then ran through。
4. Go, Tiled VAE: Replace VAE Decode with VAE Decode (Tiled), tile_size Fill 512, overlap 64。
5. Start parameters (optional): python Main. py and –reserve-vram 1.5 for system 1.5G. The old version wants to force –lowvram to add –disable-dynamic-vram –lowvram。
6. Output: Point Queue Prompt, look at the visible occupation。
Checkpoint
1 CUDA out of memory
2 Visible occupancy ratio significantly lower than fp16 (fp8/GGUF approximately 5.5 ~ 7.5 G, fp16 about 23 G - numbers estimated by third parties, not personally measured, based on your own machine)
3 also blast resolution and drop one more / –cache-none / Confirm T5 and change the quantitative version (T5 fp16 about 9G is a large household)
4 The figure is out, but it's not a problem, it's too low resolution or too few steps, and it's going back to modulation。
Job (selection): Same hint, fp16 and fp8, run side by side, feel "quantify how much paint is lost" -- - I've measured it myself. FLUX fp8 and fp16 have very little eye difference and very high value for money。
V. Core knobs: comparison of start-up parameters (printed)
These are the real start-up parameters that I verified against official documents. Add python main.py Back:
| parameter | corresponds English -ity, -ism, -ization | For what |
| –reserve-vram 1.5 | PRESERVE NGB MEMORY FOR SYSTEM | Fear and system failure will increase |
| _disable-dynamic-vram | Turn off AutoFat Management | only if you want to force the old –lowvram |
| –lowvram | Dynamic VRAM only takes effect when off | 4-8G, OLD ENVIRONMENT |
| –novram | more than-lowvram | –lowvram is not the last resort |
| –disable-smart-memoory | Radically unmount to memory | When there's a very small amount of stability on the surface |
| –cache-none | No Cache Node Results | The memory is still on the clock |
| _preview-method none | turn off/maintenance of node preview as none (no already defaulted) | If you change the preview manually, you can save that |
One sentence:every day just –reserve-vram 1.5; don't touch –lowvram expecting it to save his life (the new version is useless); it's really time-saving, then –cache-none。
VI. Common pits (all at once)
1. I'VE ONLY FORGOTTEN, T5: T5 OF FLUX ABOUT 9G IS AN INVISIBLE BOMB THAT MUST BE QUANTIFIED TOGETHER。
2. superstitious -lowvram: A new version of Dynamic VRAM is not effective on it, Bucka。
3. Resolution and visible square relationship: 20482 has a peak of about 8G, down to 5122 by a large fraction。
4. VAE didn't open, Tiled: High-level map decoders were then pulled out of the open at once, and the VAE Decode (Tiled) split。
5. fp8 in command line instead of node: –fp8_e4m3fn-unet is often invalid for FLUX loaders and is set in the Road Diffusion Model node weight_dtype。
vii. Quick check of parameters (copying)
| take | Model Format | Resolution | Key Actions |
| 4060/8G RUN FLUX | fp8 or GGF Q4_K_M | 5122-7682 Commencement | T5 also quantified + VAE Decode |
| I want to make sure it's good | fp16 (12G+card only) | 10242 cases | –reserve-vram 1.5 |
| Video visual explosion | GGF QUANTIFIED + FRAME REDUCTION | Decrease Resolution | VAE Decode (Tiled) Time Segment |
| I can't get up | GGF MINIMUM | 5122 | –cache-none +-disable-dynamic-vram –lowvram |
VIII. 4060 / 8G SPECIALIZED & I STEPPED ON THE PIT
SKILLS: 4060 THIS 8G CARD, RUNNING FLUXThe best solution is not to de-start parameters, it's to change the model format–fp8 or GGUUF Quantified version, visible directly from ~23G to 6~12G, Dynamic VRAM to help you with automatic movement, which is basically stable. Only one –reserve-vram 1.5 is added to the start parameter。
Trail experience:i once added all of the –cache-none and –lowvram brains to "extreme savings," and the picture came out, but a picture ran for more than three minutes and slowed to doubt life. later on, it became clear: cache and partitions are "provincial memory exchange speeds" and can be left unconnected. low-visibility first leverage is always model format, not these switches。