alesha-pro/tools: ComfyUI workflows and agent skills measured on 4x RTX 3090

A collection of standalone sections, not one installable package

The repository is organised as a shelf of separate projects rather than a single tool with a shared setup. The project notes that every section stands on its own, with its own installer and its own docs, and that there is no shared setup to get through first. Top level entries confirm the layout: directories for minimax-h3, qwen-image-2.1, qwen-image-2.1-gguf, and skills, alongside a README, a LICENSE, and a .gitignore. The project description frames the origin plainly, as tools, ComfyUI workflows and benchmark configs from a 4x RTX 3090 local-inference rig, and the opening line says these are things the author builds while running models on their own hardware, cleaned up enough for other people to use. The licence is MIT, and the licensing note is blunt: take it, change it, ship it. There are no GitHub releases, so you clone the repository and read the section you need.

minimax-h3: video and synchronized audio from one generation

The first section targets the MiniMax H3 open weights model, and the pitch is about a single pass. The documentation explains that MiniMax H3 writes the picture and the synchronized audio together, so voice, effects and music come out of the same generation as the frames, rather than being muxed in later by a separate tool. Thirteen ComfyUI workflows are provided in two sets. The first set runs on any single card with 24 GB of memory. The second spreads the work across every card in a multi-GPU box, and a 3.12x speedup on four cards is reported, which it frames as taking a 15-second shot from 19 minutes of sampling down to 6:41. The section also ships the custom nodes the workflow depends on, a patch described as making the turbo LoRA actually load, and a three-command installer. The self check is the first thing to run, since it tells you whether your box can host the workflows at all.

cd minimax-h3 && ./install/setup.sh check

Running it from the section directory asks the installer to verify the environment and print what is present or missing.

The multi-GPU numbers start with a power limit check

Before any of that, the documentation asks you to check one thing on your own machine, and this is the most useful paragraph in the repository. Cards ship with whatever power limit the vendor set, and it is often well under the default. On the author's machine the cards were set to 220 W against a 350 W default, and raising them to 320 W took an identical run from 220 seconds down to 170. The query below is the check, and the three fields it returns are the GPU name, the current power limit, and the default limit the card came with.

nvidia-smi --query-gpu=name,power.limit,power.default_limit --format=csv

The stated reasoning behind the whole repository is that the numbers in the section guides come from real runs on that machine rather than from datasheets, with the caveat that your absolute times will differ while the ratios usually hold. That is a fair and narrow claim, and it applies to every figure here, including the 3.12x multi-GPU result and the 19 minute to 6:41 sampling times.

qwen-image-2.1: image generation on a single 24 GB card

The second section covers image generation, editing, and transparent PNG output, and its baseline is deliberately modest: one RTX 3090 with 24 GB of VRAM, running official INT8 ConvRot weights. The section ships five native ComfyUI workflows, an isolated installer, SHA-256 model verification, and a script that submits a generation and checks the resulting PNG, which is a nicer safety net than discovering a corrupt download after a long run. Documentation comes in three flavours, an English guide, a Russian guide, and an installation brief written for an agent, which is unusual and useful if you are driving the setup through a tool that reads instructions. The section ships its own directory rather than borrowing the setup of the first one, which is the practical consequence of the no shared setup rule.

cd qwen-image-2.1
python3 install/setup.py check

The check has the same shape as the other section, one command from inside the section directory, and it reports what the machine is missing before you commit to a multi-gigabyte download.

A separate GGUF directory and a low-VRAM path for smaller cards

There is a second directory in the repository for Qwen Image 2.1 that is not mentioned in the section list, named qwen-image-2.1-gguf, and the low-VRAM guide is the entry point to that path. For cards smaller than the 24 GB baseline, the guide covers GGUF Q4_K_M quantization, CPU offloading, seven workflows, and measured memory requirements, so you can size the job against your own VRAM instead of guessing. Quantization and offloading are the two levers that matter here, and both cost time: a Q4_K_M model is smaller to hold, while CPU offloading moves part of the work across the PCIe bus, and the rig this repository is measured on is built on PCIe 3.0 x16 with no NVLink, so the bus bandwidth is part of every number involved. The guide gives measured memory requirements rather than a vague warning, which is the difference between a document you can plan with and one you cannot.

skills: canvas animation and product films without a video model

The skills directory holds agent skills that also stand on their own, and they are the most surprising part of the repository because they need no GPU at all. The hand-drawn-canvas-animation skill produces short films with authored poses and expressive strokes in JavaScript and Canvas 2D, covering pencil, ink, risograph, screen print, and drawings interacting with real photos, plus sand animation and projected paper. A 60-second phoenix film in the examples folder combines those styles through a single story, and the skill includes editable examples, original sound, an MP4 renderer, and browser regression checks, with no Blender and no video generation model required. Installing it is a copy into the agent skills directory.

cp -r skills/hand-drawn-canvas-animation ~/.agents/skills/

The motion-reel skill is the commercial counterpart, turning a brief and visual references into product films, showreels, and UI morphs, with a Canvas renderer, spring motion, original music and SFX, beat analysis, contact sheets, and MP4 checks. Both cover planning, animation, review, and export, and both ship source files you can edit.

What this repository is not

The limitations are easy to miss and worth stating. There is no packaging, no release artefacts, and no shared dependency management, so a version mismatch between two sections is possible and nothing in the repository checks for one. Every performance figure comes from a single consumer rig of four RTX 3090 cards with 96 GB of VRAM in total, 320 W per card, PCIe 3.0 x16, no NVLink, running Ubuntu, and although the caveat is printed up front, the numbers are still quoted as headline results elsewhere. The repository is a personal toolbox that happens to be public, and the closing line, that more lands here as it is cleaned up, sets the expectation that sections appear over time rather than a stable set of tools. The documentation also does not cover a rollback, a compatibility matrix, or a minimum driver version for the custom nodes. If you need a maintained distribution with versioned releases, or a project you can file issues against with a maintainer who will answer, this is the wrong shape of dependency.

Editorial conclusion

Adopt alesha-pro/tools if you run ComfyUI on NVIDIA hardware and want working MiniMax H3 or Qwen Image 2.1 workflows with the custom nodes and the LoRA loading patch already handled, and if you can accept that every number in the repository was measured on one four card 3090 box. Do not treat it as a general benchmark suite or as maintained software: the repository has no releases, and each section states its own hardware assumptions. Before running anything, check the power limit on your own cards with nvidia-smi, because the first warning in the documentation is that a stock power limit can cost you around a quarter of your generation time.

Frequently asked questions

Official sources

1. alesha-pro/tools on GitHub
2. Issues
3. License: MIT
4. README