15
Powered by @reddit2telegram and @r_channels
MiniMax H3 single frame edit excels at character consistency and identity preservation. Do we really need character LoRAs anymore?
https://redd.it/1xqx0jt
@stablediffusion_r
Some interesting result when test styles on Kroma
https://redd.it/1xpw26m
@stablediffusion_r
My mom described her dream house. And I tried to make it😭
https://redd.it/1xq2q1q
@stablediffusion_r
SCAIL-2 vs MiniMax for Motion Control – RTX 5090
I'm using an RTX 5090 (32 GB VRAM) on RunPod for AI motion-control videos.
I'm currently testing SCAIL-2 in ComfyUI. A 5-second video at 720p takes around 25 minutes to generate.
Is SCAIL-2 or MiniMax better for motion control in terms of quality, speed and cost? Which would you recommend for realistic character movements while keeping GPU rental costs low?
Could anyone share their ComfyUI workflows or JSON workflow files for motion control using an RTX 5090?
Any tips or workflow recommendations would be appreciated!
https://redd.it/1xq2wze
@stablediffusion_r
p-h3-image 2 step lora is out
https://redd.it/1xq1po9
@stablediffusion_r
viggle image swap for multi character swap
https://redd.it/1xpyeta
@stablediffusion_r
它是否被严重低估了？
https://redd.it/1xpxs3y
@stablediffusion_r
First test of Geiru Context Suite to produce a 2 minute shot! Works great!!
https://redd.it/1xpuf8b
@stablediffusion_r
Watch or capture live video+audio or images from another browser tab right in ComfyUI
https://redd.it/1xporz5
@stablediffusion_r
Dario Amodei: A lot of AI generated videos and images are fake, addictive, and just slop.
https://redd.it/1xpqmye
@stablediffusion_r
Changing a character across an ~hour long TV episode
https://redd.it/1xpo5g6
@stablediffusion_r
Testing the new RelayVSR vs SeedVR2 vs native finish for 1080P
https://redd.it/1xph4o1
@stablediffusion_r
With each tile measuring 1344 × 768, a single 4K frame is physically divided into a 4 × 4 grid = 16 tiles.
* Those 16 sampler runs, each lasting 1 minute and 7 seconds, were the 16 tiles in the first window.
* Window 1/7 took exactly: 16 tiles × 1.1 minutes ≈ 18 minutes of sampling time.
2. The final math for 10 seconds in 4K:
* 7 windows × 16 tiles = 112 sampler passes.
* 7 × 18 minutes ≈ 2 hours.
* So generating 10 seconds of video in true 4K (243 frames) on a consumer 16 GB GPU will take about 2 hours in total. The upside is that it should be completely stable and avoid out-of-memory errors
https://redd.it/1xph0pt
@stablediffusion_r
ComfyUI-MiniMaxH3-TemporalWindow-SpatialTiles works on 16GB VRAM + 32GB RAM with NO OOM
I modified some workflow and added to it ComfyUI-MiniMaxH3-TemporalWindow-SpatialTiles (from this post [https://www.reddit.com/r/StableDiffusion/comments/1xgcjvw/i\_created\_a\_custom\_node\_that\_can\_help\_you/](https://www.reddit.com/r/StableDiffusion/comments/1xgcjvw/i_created_a_custom_node_that_can_help_you/) ).
My 10sec video with **960x544** resolution could take 2 hours for 4K (3840×2176).
No OOM at all !!!
Workflow itself : [https://pastebin.com/xJpm84f7](https://pastebin.com/xJpm84f7) 😍
`[MinimaxH3-3D] Latent 60x34 -> 240x136 | Pixels 3840x2176 | scale=4.000`
`[MinimaxH3-3D] temporal chunking: T=72 chunks=3 overlap=5`
`[MinimaxH3-3D] ✅ Model offloaded to CPU. VRAM released.`
`[INFO] Found quantization metadata version 1`
`[INFO] Detected mixed precision quantization`
`[INFO] Using mixed precision operations`
`[INFO] Native ops: convrot_w4a4, float8_e4m3fn, int8_tensorwise, float8_e5m2, asym_w4a8_int8, w6a8_int8 , emulated ops: mxfp8, nvfp4`
`[INFO] model weight dtype torch.bfloat16, manual cast: torch.bfloat16`
`[INFO] model_type FLOW_AV`
`[INFO] Model storage policy: fast_disk=True paths=['D:\\ai\\ComfyUI_windows_portable_nvidia\\ComfyUI_windows_portable\\ComfyUI\\models\\diffusion_models\\minimax_h3_ref2va_pruned_int8_convrot.safetensors']`
`[INFO] MiniMax H3 windows: 7 window(s) x 16 tile(s), 243 frames (CLIP not routed through the node: every window reads the full-shot prompt).`
`[INFO] Requested to load MiniMaxH3`
`[INFO] Model MiniMaxH3 prepared for dynamic VRAM loading. 19995MB Staged. 208 patches attached. Force pre-loaded 210 weights: 1175 KB.`
`100%|████████████████████████████████████████████████████████████████████████████████████| 2/2 [02:09<00:00, 64.68s/it]`
`[INFO] 0 models unloaded.`
`[INFO] Model MiniMaxH3 prepared for dynamic VRAM loading. 19995MB Staged. 208 patches attached. Force pre-loaded 210 weights: 1175 KB.`
`100%|████████████████████████████████████████████████████████████████████████████████████| 2/2 [01:09<00:00, 34.93s/it]`
`[INFO] 0 models unloaded.`
`[INFO] Model MiniMaxH3 prepared for dynamic VRAM loading. 19995MB Staged. 208 patches attached. Force pre-loaded 210 weights: 1175 KB.`
`100%|████████████████████████████████████████████████████████████████████████████████████| 2/2 [03:14<00:00, 97.15s/it]`
`[INFO] 0 models unloaded.`
`[INFO] Model MiniMaxH3 prepared for dynamic VRAM loading. 19995MB Staged. 208 patches attached. Force pre-loaded 210 weights: 1175 KB.`
`100%|████████████████████████████████████████████████████████████████████████████████████| 2/2 [01:08<00:00, 34.12s/it]`
`[INFO] 0 models unloaded.`
`[INFO] Model MiniMaxH3 prepared for dynamic VRAM loading. 19995MB Staged. 208 patches attached. Force pre-loaded 210 weights: 1175 KB.`
`100%|████████████████████████████████████████████████████████████████████████████████████| 2/2 [01:09<00:00, 34.87s/it]`
`[INFO] 0 models unloaded.`
`[INFO] Model MiniMaxH3 prepared for dynamic VRAM loading. 19995MB Staged. 208 patches attached. Force pre-loaded 210 weights: 1175 KB.`
`100%|████████████████████████████████████████████████████████████████████████████████████| 2/2 [01:09<00:00, 34.51s/it]`
`[INFO] 0 models unloaded.`
`[INFO] Model MiniMaxH3 prepared for dynamic VRAM loading. 19995MB Staged. 208 patches attached. Force pre-loaded 210 weights: 1175 KB.`
`100%|████████████████████████████████████████████████████████████████████████████████████| 2/2 [01:07<00:00, 33.81s/it]`
`[INFO] 0 models unloaded.`
`[INFO] Model MiniMaxH3 prepared for dynamic VRAM loading. 19995MB Staged. 208 patches attached. Force pre-loaded 210 weights: 1175 KB.`
`100%|████████████████████████████████████████████████████████████████████████████████████| 2/2 [01:07<00:00, 33.80s/it]`
`[INFO] 0 models unloaded.`
`[INFO] Model MiniMaxH3 prepared for dynamic VRAM loading. 19995MB Staged. 208 patches attached. Force pre-loaded 210 weights: 1175
How are you fixing the lighting shift when doing continuation videos with Minimax?
Im using the same seed, same sampler settings and same prompt for the lighting but I always get a slight color/lighting shift.
The only thing i could think of is a color match node but ive never had much success with them and i dont think i have enough ram for large videos.
Im using the seed hunter workflow. With the Singularity model and DMAD turbo lora.
https://civitai.red/models/2881362/minimax-h3-seed-hunter-latent-upscaler-seamless-video-continuation-dialogue-gen-tool-speedups-ref-mods?modelVersionId=3453566
https://redd.it/1xpdsyd
@stablediffusion_r
Microsoft engineer built a tool that shows what's really using your GPU
https://www.neowin.net/news/microsoft-engineer-built-a-tool-that-shows-whats-really-using-your-gpu/
https://redd.it/1xq94xa
@stablediffusion_r
Draw Things Mac – Help needed with custom.json model entries
I'm trying to restore some existing Draw Things models from an external HD without having to download them all again.
I've discovered that my custom.json file is missing the registration entries for these models:
1. FLUX.1 Schnell Q5 – flux_1_schnell_q5p.ckpt
2. ERNIE Image Turbo i8x – ernie_image_turbo_i8x.ckpt
3. Qwen 3.5 4B i8x – qwen_3.5_4b_i8x.ckpt
I still have the original .ckpt and matching -tensordata files.
Would anyone with these models working in Draw Things on Mac be kind enough to copy and paste the corresponding entries from their custom.json file?
I'm hoping to reconstruct the missing JSON entries and get my existing models working again without downloading several gigabytes of files.
If successful, I'll share the recovery process with the community and make a video to help others in the same situation.
Many thanks for any help!
https://redd.it/1xq4rl5
@stablediffusion_r
TBG ComfyUi Node for Layered Decomposition – RevealLayer
https://redd.it/1xq27mf
@stablediffusion_r
C++ Desktop Client for ComfyUI
When I first started with ComfyUI, the first thing I looked for was a native desktop client without browser lag and DOM-based architecture.
I'm a C++ dev with over 20 years of commercial experience, so I eventually just started writing my own client on the side. I used an ultra-fast UI framework that can easily chew through thousands of nodes on a canvas without the slightest stutter, boost for networking, sqlite as a DB, and the fmt library - the fastest formatting framework out there, leaving room for an easy port to Linux/Mac.
The main reason all existing clients are forced into the browser is custom JS nodes, which are themselves bottlenecked by Litegraph's architecture. Bypassing this turned out to be easier than I thought. The internal logic of these nodes is actually quite simple, so I ported most of it to C++. I also built extended versions of frontend-only nodes like reroutes and bookmarks.
For the parts that had to stay in HTML/JavaScript, I isolated them in a completely separate process. This ensures the heavy browser engine never interferes with the lightweight C++ core.
https://preview.redd.it/6hd5unvqj50i1.png?width=1938&format=png&auto=webp&s=02fe9e319c2fab0f5ab6af726a5c4b233c4ce2d5
I recently finished a stable Beta. Check it out, the link to the source code and prebuilt Windows binary is in the comments. I also added a demo video (making a video was another cool new experience for me).
https://redd.it/1xq2eei
@stablediffusion_r
Updated hyperflow-h3-flash (4, 6, or 8 steps)
https://huggingface.co/VideoRebirthAI/hyperflow-h3-flash
https://redd.it/1xq0kuk
@stablediffusion_r
Rebuilt Higgsfield's "Miracle" demo edit by edit with Viggle-remake
https://redd.it/1xpz5ti
@stablediffusion_r
Why are all the newly released Image Edit models junk?
I was trying to test the new Viggle-Remake-Swap LoRA in MiniMax-H3. For this I needed to character swap in one of the frames of my source video clip (3.75s, 90 frames), so that the target video would have the new character in it. I went through all the new image edit models released recently to do the character swap. I tried Flux.3 Klein, Qwen 2.1, Fibo Image Edit 2.0, p-H3-Image-Edit and they all failed. So, I went back to the older image edit models Qwen 2511, Klein 9B, Krea2-Identity-Edit v1.2. The first two worked perfectly, while Krea2 Identity Edit did a far better job than any of the newer edit models. Final conclusion - Flux.3 Klein: crap much much worse than even Klein 4B; Qwen 2.1: really crap (does a cut-and-paste job); Fibo 2.0: really crap; p-H3-Edit: absolute junk, doesn't edit anything, input and output are the same; Qwen 2511: still amazing, slight issue with facial blur; Klein 9B: absolutely perfect and still the fastest; Krea2-Identity-Edit: very good edit job, but didn't follow the pose change part of the prompt.
While the video and T2I models in the OS domain seem to be improving, why are the I2I or Image Edit models getting worse? And more importantly, why are people eagerly using those newer models like salivating dogs and posting rave reviews about them? As I see it, all those newly released Image Edit models are junk and need to be kicked to the curb so that their developers may learn some lessons, hopefully.
https://redd.it/1xpwnro
@stablediffusion_r
Moss SFT INT8-convrot is just fine
https://redd.it/1xpu1e0
@stablediffusion_r
The Problem with Jean-Paul Junior
https://redd.it/1xpr1ec
@stablediffusion_r
Has anyone managed to get an H3 image edit workflow working on a Mac?
I currently use DrawThings as my "daily driver" with MiniMax H3 but I don't see it getting support any time soon. Does anyone know of any workflow in Comfy or elsewhere that are somewhat out of the box. Thanks!
https://redd.it/1xpm843
@stablediffusion_r
Made a face swap that runs entirely in your browser, no setup or huge downloads
https://redd.it/1xpkwtr
@stablediffusion_r
I hosted a live San Altman avatar locally on my RAM
https://redd.it/1xpiuyk
@stablediffusion_r
KB.`
`100%|████████████████████████████████████████████████████████████████████████████████████| 2/2 [01:07<00:00, 33.79s/it]`
`[INFO] 0 models unloaded.`
`[INFO] Model MiniMaxH3 prepared for dynamic VRAM loading. 19995MB Staged. 208 patches attached. Force pre-loaded 210 weights: 1175 KB.`
`100%|████████████████████████████████████████████████████████████████████████████████████| 2/2 [01:08<00:00, 34.05s/it]`
`[INFO] 0 models unloaded.`
`[INFO] Model MiniMaxH3 prepared for dynamic VRAM loading. 19995MB Staged. 208 patches attached. Force pre-loaded 210 weights: 1175 KB.`
`100%|████████████████████████████████████████████████████████████████████████████████████| 2/2 [01:09<00:00, 34.73s/it]`
`[INFO] 0 models unloaded.`
`[INFO] Model MiniMaxH3 prepared for dynamic VRAM loading. 19995MB Staged. 208 patches attached. Force pre-loaded 210 weights: 1175 KB.`
`100%|████████████████████████████████████████████████████████████████████████████████████| 2/2 [01:07<00:00, 33.63s/it]`
`[INFO] 0 models unloaded.`
`[INFO] Model MiniMaxH3 prepared for dynamic VRAM loading. 19995MB Staged. 208 patches attached. Force pre-loaded 210 weights: 1175 KB.`
`100%|████████████████████████████████████████████████████████████████████████████████████| 2/2 [01:07<00:00, 33.74s/it]`
`[INFO] 0 models unloaded.`
`[INFO] Model MiniMaxH3 prepared for dynamic VRAM loading. 19995MB Staged. 208 patches attached. Force pre-loaded 210 weights: 1175 KB.`
`100%|█████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 2/2 [01:06<00:00, 33.46s/it]`
`[INFO] 0 models unloaded.`
`[INFO] Model MiniMaxH3 prepared for dynamic VRAM loading. 19995MB Staged. 208 patches attached. Force pre-loaded 210 weights: 1175 KB.`
`100%|█████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 2/2 [01:07<00:00, 33.67s/it]`
`[INFO] 0 models unloaded.`
`[INFO] Model MiniMaxH3 prepared for dynamic VRAM loading. 19995MB Staged. 208 patches attached. Force pre-loaded 210 weights: 1175 KB.`
`100%|█████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 2/2 [01:07<00:00, 33.75s/it]`
`[INFO] MiniMax H3 windows: window 1/7 done (frames 0-55, 0 frames carried in).`
`[INFO] 0 models unloaded.`
`[INFO] Model MiniMaxH3 prepared for dynamic VRAM loading. 19995MB Staged. 208 patches attached. Force pre-loaded 210 weights: 1175 KB.`
`100%|█████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 2/2 [01:33<00:00, 46.64s/it]`
`[INFO] 0 models unloaded.`
`[INFO] Model MiniMaxH3 prepared for dynamic VRAM loading. 19995MB Staged. 208 patches attached. Force pre-loaded 210 weights: 1175 KB.`
`100%|█████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 2/2 [01:33<00:00, 46.57s/it]`
`[INFO] 0 models unloaded.`
`[INFO] Model MiniMaxH3 prepared for dynamic VRAM loading. 19995MB Staged. 208 patches attached. Force pre-loaded 210 weights: 1175 KB.`
`0%| | 0/2 [00:00<?, ?it/s, Model Initializing ... ][INFO] Interrupting prompt 58585033-a8c5-4dcb-8004-6e0e6de9786c`
`0%| | 0/2 [00:36<?, ?it/s, Model Initializing ... ]`
`[INFO] Comfy model compiler graph breaks: 3, rogues: 10`
`[INFO] Processing interrupted`
^(What this log says) :
# What’s actually happening:
1. 16 tiles per window:
* Because you set `scale: 4.00`, the resolution became 4K (3840 × 2176).
*
Current best H3 speed-up techniques?
I used H3 heavily the first two weeks and a ton of amazing improvements for improving speed at small quality trade-off came out (Sparse Attention comes to mind).
I imagine things have matured a lot since then and probably with new approaches since then!
I'm on a 4090 so that's my target, but I'm curious for all tiers.
(I couldn't find a thread of recent times discussing this, if there is one, I apologize and I'll remove my post.)
https://redd.it/1xpfrx3
@stablediffusion_r
Mini max h3 animation workflows
https://redd.it/1xp792y
@stablediffusion_r
