ComfyUI on P104-100: the photos worked, the video needs another try

I installed ComfyUI on a server with two 8 GB P104-100 GPUs and tried a few configurations for Korean hanbok portraits. I generated four realistic images, converted one of them into three anime-style variations, and then attempted a short motion clip. The still images turned out well enough to compare. The video did not, so I am leaving it as a test rather than a finished piece.

Server and installation

An LLM evaluation was running at first, so I limited the initial work to installation and model downloads. I ran the GPU and image tests only after that evaluation ended and I had permission to proceed. ComfyUI lived in a separate Python environment and listened on the server’s localhost interface; I reached it through an SSH tunnel.

Item	Configuration used in this test
OS	Ubuntu 26.04 LTS; kernel 7.0.0-38-generic
GPUs	Two P104-100 cards, 8 GB each; generation used GPU 0 only
NVIDIA driver	580.178.04
Python / PyTorch	3.12.15 in a separate environment / 2.8.0+cu126
CUDA build	12.6, chosen with Pascal compatibility in mind
ComfyUI commit	c06bf5f569d268b80076e5f44126f87cdf835869
Image-generation mode	FP32, normal VRAM, split cross-attention
Access	localhost:8188 on the server, via SSH tunnel

I tested simple FP32 matrix operations on each card, but I did not pool their VRAM. The installation took a couple of detours: a CUDA dependency download timed out, and the first ComfyUI launch stopped because --normalvram was not supported in this setup. After adjusting the download path, I relaunched with --disable-dynamic-vram.

Checkpoints, LoRAs, and node connections

For realistic portraits I used the majicMIX Realistic v7 checkpoint; for the anime-style conversions I used MeinaMix v11. Hanbok V2 is a LoRA, not a checkpoint. In ComfyUI, I fed the checkpoint’s MODEL and CLIP outputs into Load LoRA, sent the LoRA MODEL output to KSampler and its CLIP output to the prompt encoders, and used the checkpoint’s VAE. I chained another Load LoRA node when testing a second LoRA. I verified the basic SD 1.5 checkpoint file, but did not use it for these images.

Purpose	Model and LoRA combination	Strength in this test
Realistic baseline	majicMIX Realistic v7	No LoRA
Hanbok	majicMIX + Hanbok V2	0.65
Portrait variation	majicMIX + Hanbok + Korean Doll Likeness v2	0.65 + 0.4
Film variation	majicMIX + Hanbok + FilmGirl v4	0.65 + 0.4
Anime img2img	MeinaMix v11 + Hanbok V2	Hanbok 0.65

Four realistic hanbok portraits

I kept the prompt, seed 20261018, resolution 512 × 768, 20 steps, Euler/normal sampler and scheduler, and CFG 6.5 fixed. Each condition was run once. These are wall-clock times from submitting an API request to its completed history entry; the first run includes checkpoint loading.

Condition	API request time
Baseline	51.77 s (includes initial checkpoint load)
Hanbok 0.65	54.48 s
Hanbok 0.65 + Korean Doll 0.4	50.07 s
Hanbok 0.65 + FilmGirl 0.4	50.07 s

The four realistic results. Only the LoRA combination changes between conditions; this is not a test of historical costume accuracy.

API wall-clock time for each realistic-image request. Each condition was measured once.

With Hanbok V2, the shorter jacket and ribbon-like detail became more noticeable than in the baseline. Adding the portrait LoRA changed the face and eyes as well. The FilmGirl result changed some colors and collar details, but I could not clearly pick out a film-grain effect. I would not treat four single runs as a general performance benchmark.

Turning a portrait into anime art

Next I used one generated hanbok portrait as the input to MeinaMix img2img. I held seed 20261018, 20 steps, DPM++ 2M/Karras, CFG 7, CLIP skip 2 (last layer -2), and Hanbok strength 0.65 steady. Only denoise changed.

Denoise	API request time	What I saw in this input
0.35	41.06 s (includes first MeinaMix load)	Much of the photographic texture remains
0.55	19.24 s	Photographic and drawn qualities mix
0.75	19.21 s	The clearest anime-style linework and flat shading

The source portrait beside the 0.35, 0.55 and 0.75 denoise results. I selected 0.75 as the reference for the video test.

API wall-clock times for the three img2img requests. The first includes loading MeinaMix.

For this particular photo, 0.75 looked most like an anime illustration to me. That came with a trade-off: the face and clothing details were not preserved perfectly. The first request loaded a checkpoint, so these timings do not establish that one denoise setting is inherently slower.

The one-second video did not work out

I tried to place that anime-style character in a full-body beach scene and animate an index-finger “no” gesture. The workflow used AnimateDiff, IP-Adapter Plus, OpenPose and Advanced-ControlNet. One run produced a 640 × 360 video with eight frames at 8 fps. CPU FFmpeg then resized it with Lanczos and interpolated it to a 3840 × 2160, 24 fps output file.

Step	Measured time or status	What it means
Successful one-second generation	268.236 s	API submission to completion
Output-format conversion	5.057 s	CPU resize and interpolation; no AI super-resolution
Successful path total	273.293 s (about 4 min 33 s)	Excludes installation and failed retries
Ten-second final video	Not run	Stopped after inspecting the one-second test

All eight source frames. The hands, face and background are inconsistent, so I did not use the clip as a publishable video.

The MP4 was created and decoded successfully, but the background became blotchy, the character changed between frames, and the intended finger gesture was hard to read. The “4K” version describes the output dimensions, not native 4K generation or recovered detail. I did not publish the motion clip or run the planned ten-second version. A separate twelve-second slideshow was assembled from still images; it is not AI-generated motion footage.

What I would try next

Image generation and the hanbok-to-anime comparison worked on this server. The video experiment showed me that producing a valid file is not the same as producing a usable clip. I stopped the GPU work here. I may revisit video later with different settings or another approach, but this run does not tell me how well a longer clip would work.

The Hanbok V2 LoRA was created by MassBrainImpact. Here is the original model page.