Next hour · @devraj0390
49 followers. The title plus #comfy #ltx #localai plus a YouTube link will not leave the account. The useful version is sitting under it as a reply, with a placeholder chapter, where nobody will see it.
Latest uploads are stalling at 174, 209, and 216 views, with 0–2 comments. 62.6 hours across 8.2k views is about 27 seconds each. That is a shorts channel that keeps uploading 40-minute videos. The thumbnail on this one asks seven questions, including a question mark and a coffee banner. One claim next time, no question mark: 16GB. It ran.
X buries outbound links. The URL goes in the reply.
32GB on the box. 16GB on the desk. LTX-2.5 finished on a Mac Mini M4. Not an install video. The ComfyUI graph that stops the MPS crash: • Video VAE and audio VAE stay the stock files • Gemma 4 12B stays INT8. Rename the CLIP and this build dies • The 22B runs as GGUF, not pinned whole • 740×480 on 16GB, then the upscaler • FP4 and FP8 fight this machine Reply with your Mac and the RAM.
The full cut, generation included: https://youtu.be/OuxRE4m8PU8 Written version, no paywall: https://medium.com/@aiwithjsdev/32gb-on-the-box-16gb-on-the-desk-5bc7cb731330
Do not open with “watch my video.” The fact is the reply. Leave the link for when they ask.
LTX 2.5 on an RTX 5060 Ti 16GB · same wall, NVIDIA side
Same 16GB wall, different backend. On the M4 the graph that held was stock video VAE + audio VAE, Gemma 4 INT8 with the CLIP name left alone, and the 22B as GGUF at 740×480. FP8 is what throws invalid buffer size on MPS. NVFP4 can stay on the 5060 Ti. The Mini cannot load that file whole.
LTX-2.5 in one Comfy node · the graph people are saving
That single node is the right UI. On a 16GB M4 it still dies if Gemma INT8 and the full transformer are resident together, or if the CLIP gets renamed. Gemma runs, then leaves. GGUF the 22B. Tile 740×480.
AI WITH JS DEV · Mac Mini M4 · ComfyUI
Gemma INT8 is allowed to run. It is not allowed to stay while the full transformer loads. Encode, then sample. Below is that map, then the rest of the posts.
Peak
18.6GB
Holds
2.6GB into swap while the big file is resident. Under the 21GB wall. The video finishes this.
Encode is the heavy moment. Gemma INT8 is 15.4GB and then it leaves. The sample is the smaller number.
18.6
13.8
Model of the video, not a profiler. GGUF weights are counted as partly resident. Audio VAE is an estimate (0.7GB); video VAE is the published 1.5GB file. Official sizes are the public LTX-2.5 downloads.
The 16GB tile from the video.
Why the download misses
Before a frame35.6 GB
Mac Mini M416 GB
Apple silicon has no spare VRAM chip. The same pool is the GPU, the CPU, and macOS. Load both giants and the Mini is already in swap before ComfyUI samples anything.
How the cut works
01
Gemma runs, then it leaves. The video VAE and the transformer show up for the sample. Peak is the heavier phase, not the sum of the files.
02
Gemma 4 12B stays the INT8 file. Do not rename the CLIP. The 22B distilled transformer is the one that becomes GGUF, so it is not pinned whole. FP4 and FP8 are what fight MPS.
03
This LTX-2.5 build crashes when the clip name changes. Use a GGUF clip loader and leave the name it expects. That was the silent killer, not the sampler.
04
Video VAE and audio VAE stay the stock files. The quant is for the 22B, not the decoders.
05
That is the frame size the video gives a 16GB machine. 1200px is a spatial upscaler pass on top, not a native render. 4K is the 32GB conversation.
06
On 16GB, stay near 14GB so macOS still exists. If you spill into swap, stop before 21GB of it. Past that the job does not come back.
When it dies
01
Gemma INT8 and a full transformer will not sit down together. GGUF the 22B, tile 740×480, let Gemma leave before the sample, and close everything else.
02
The MPS failure the video hits with FP8 and FP4, and with the transformer’s own INT8 — not with Gemma’s. F16 holds. GGUF Q4 and Q8 hold. The creator was taking that bug to PyTorch.
03
Expected. This build is brittle about the text-encoder name. Put the GGUF clip loader in and stop renaming the node to something tidy.
04
Swap in the GGUF LoRA loader — the 2.3-style one — instead of the stock loader. If you are on FP4, that quant is the other half of the crash.
05
Watch swap, not the spinner. The line in the video is 21GB of swap on a 16GB system. Cross it and you are not waiting. You are done.
The 40 minutes
Chapter marks follow the upload. Workflow files are the creator’s, not a free mirror — Buy Me a Coffee. Related: the earlier test and the crash fix, linked from that description.
For the channel
The post for tonight is at the top. This is the rest of the desk: titles, three shorts, the pin, the description, the thread. Your last three videos died quietly because nobody had a one-line thing to answer.
First line of a short, or the first comment under a duet.
LTX-2.5 asks for about 32GB. This 16GB Mac Mini just finished a clip.
The M4 does not die on Gemma INT8. It dies if you rename the CLIP, or if FP8 shows up.
Your workflow is fine. You renamed the clip. This build crashes when you do that.
740×480 is not a sad resolution. It is the tile that stays under 14GB.
Swap past 21GB and the Mini does not slow down. It stops coming back.
One job each. Contradiction outperforms a feature list at this size.
Contradiction
LTX-2.5 on a 16GB Mac Mini M4: The ComfyUI Cut That Actually Holds
Search
LTX-2.5 ComfyUI on Mac Mini M4 16GB — GGUF Settings, 740×480, No MPS OOM
Specific
I Kept LTX-2.5 Under 14GB on a 16GB M4. The Clip Name Was the Crash.
Stakes
Official LTX-2.5 Is a 36GB Download. This 16GB Mini Never Loads It That Way.
Do not reshoot. Cut the long video at these beats.
18s · Cut the math from the quant chapter. No intro.
ON SCREEN, huge: 18.7 + 15.4 + 1.5 VOICE: That is the official LTX-2.5 download. Before a single frame. ON SCREEN: Your Mini has 16. VOICE: So they never sit down together. Gemma INT8 runs, then it leaves. The 22B is GGUF. Tile is 740 by 480. END CARD: The whole memory map is on the channel. AI WITH JS DEV.
15s · Architecture chapter, the moment the crash is named.
ON SCREEN: It crashed when I renamed the clip. VOICE: LTX-2.5 on this Comfy build dies if you change the clip name. Not the sampler. The name. ON SCREEN: GGUF clip loader. Leave it. VOICE: That was the fix hiding inside a 40 minute video. Mac model and RAM in the comments.
16s · RAM chapter, Activity Monitor if you showed it.
ON SCREEN: 14GB ceiling. 21GB swap wall. VOICE: On a 16GB M4, LTX-2.5 is allowed to breathe up to about 14 gigs. If it falls into swap, I kill it before 21. ON SCREEN: Past 21, it does not finish. VOICE: 1200 pixels is the upscaler. It is not a native 4K render. Full cut on the channel.
The pin is the settings. The description is the search.
Drop your Mac and the RAM. 16GB or more? The cut in the video: • Gemma 4 12B stays INT8. Do not rename the CLIP or this build crashes. • The 22B distilled transformer is GGUF, not loaded whole. • Video VAE and audio VAE stay the stock files. • 740×480. 1200px is the upscaler, not the native tile. • Encode can spill a few GB. If swap crosses 21GB, stop. It will not come back. FP4 and FP8 are the invalid-buffer crashes on MPS.
LTX-2.5 is a 22B video model plus a Gemma 4 12B text encoder. The comfortable download is a 32GB-class machine. This is the ComfyUI cut that finished on a 16GB Mac Mini M4. Will your Mac hold it? Pick your RAM, the quant, and the tile — encode phase and sample phase are separate, because the trick is to never have both giants resident. The 16GB cut, short version • Gemma 4 12B stays INT8. Do not rename the CLIP or this build crashes. • Distilled transformer as GGUF (Q3 or Q4), not the official NVFP4 or transformer INT8. • Video VAE and audio VAE stay the stock files. • 740×480 native. 1200px is a spatial upscaler pass. 4K is the 32GB conversation. • Encode may spill a few GB. Swap wall is 21GB. FP4 and FP8 fight MPS. Chapters 0:00 The Mini and the claim 2:15 Why 16GB loses 5:30 ComfyUI and the MPS flags 10:15 What gets quantized 16:00 The 16GB architecture 23:50 A real generation 28:20 RAM, not vibes 34:15 When it OOMs Comment your Mac model and RAM. If you are crashing, say whether it dies at load, at the clip, or halfway through the sample. Written version: https://medium.com/@aiwithjsdev/32gb-on-the-box-16gb-on-the-desk-5bc7cb731330 #LTX25 #ComfyUI #MacMiniM4 #AppleSilicon #LocalAI
Five posts. The last one carries the video. Do not put the link in post one.
01
LTX-2.5 on a 16GB Mac Mini M4. Not a cloud box. ComfyUI, local, and it finished. The official files are the trap. Roughly 18.7GB NVFP4 transformer + 15.4GB INT8 Gemma + the VAE. That is ~36GB before a frame. The Mini has 16.
02
The cut is sequencing, not a miracle quant. Gemma 4 stays INT8, then gets out of the way. The 22B is GGUF, streamed, not pinned whole. Video VAE and audio VAE stay the stock files. Peak is the heavier phase, not the sum.
03
Two things that actually crash it, both boring: 1. Renaming the clip. This LTX-2.5 build dies when the clip name changes. GGUF clip loader, name left alone. 2. FP8 or FP4 on MPS. Invalid buffer size. The transformer’s own INT8 does it too. Gemma’s INT8 is not that bug. F16 and GGUF held. FP4 LoRAs need the GGUF LoRA loader.
04
Tile is 740×480 on 16GB. That is the number. 1200px in the video is the spatial upscaler, not a native render. 4K is the 32GB conversation. Working line ~14GB so macOS still exists. If swap crosses 21GB, kill it. It will not come back.
05
Full architecture, the live gens, and the OOM cases are in the video. 40 minutes, no installer tour. Comment your Mac + RAM if you want the settings aimed at your box. https://youtu.be/OuxRE4m8PU8
The comments that already exist are the distribution. Answer with the setting, not a teaser.
They say
16GB M4, dies as soon as I queue
You say
Gemma 4 INT8 is supposed to run first and then leave. If the full transformer is still trying to sit down with it, or you are on FP8, that is the OOM. GGUF the 22B, tile 740×480, and do not rename the CLIP.
They say
It worked until I renamed the text encoder
You say
That rename is the crash. This LTX-2.5 build wants the clip name left as it expects. GGUF clip loader, then stop tidying the node.
They say
I have 24GB. Can I leave 740?
You say
Run the 16GB cut first so you know it holds. Then try the spatial upscaler — that is how the video gets toward 1200px. Native 1080 and 4K are still the 32GB side of the map.
They say
How many minutes per clip?
You say
Depends on the tile and the quant more than the chip slogan. The timed runs are in the generation chapter, around 23:50. I would rather you match the memory than chase a number from a different resolution.
Holds
18.6GB peak · 16GB · 740×480