Перестань бояться ComfyUI. Начни локально и бесплатно

Concept: “Aggregators” and why ComfyUI is common under the hood

The video explains how to start using ComfyUI locally and for free, why people get scared by “model aggregators,” and provides a step-by-step tutorial for installing, running, and improving image-generation workflows.

What “aggregators” are

• The speaker references multiple aggregator sites (e.g., “neuroaggregators”), which:
  • Bundle many different neural networks behind a single subscription
  • Differ by:
    • interface
    • limitations
    • capabilities

Why ComfyUI comes up so often

• ComfyUI is framed as the “core” system underlying many aggregator-style offerings (“dad of all aggregators”).
• The learning curve is attributed to needing to understand many models and workflows.
• Key reassurance: once you grasp the local setup, ComfyUI becomes manageable.

---

Installation guide (local vs. cloud)

Local (free)

• In 2026, download from conf.org → Download Local (Windows).
• Installation is described as very quick (about 10 seconds for the program).
• First run requires choosing:
  • Language (English preferred)
  • GPU (select your video card)
• Notes:
  • ComfyUI can be a VRAM “glutton”
  • NVIDIA works better
  • A more powerful GPU is better
• You set a folder path for projects/models/outputs, ensuring enough free disk space.
• There’s also an option involving an unchecked “Usage” checkbox (details unclear in the summary).

Cloud (easier but paid)

• If your GPU isn’t strong enough, use Launch Cloud.
• Requires:
  • Google registration
  • a subscription
• Runs remotely but keeps a similar interface.

---

First workflow: nodes, errors, and model downloading

How workflows are represented

• The ComfyUI interface displays workflows as nodes on a canvas.
• Example text-to-image workflow nodes include:
  • a model/text encoding node
  • a save image node

Common issue: red errors before models are downloaded

• Red error indicators appear until required models are downloaded.

Debugging and downloading missing assets

• Use “Show missing models”
• Download the required models/assets
  • Example total size cited: roughly ~19.5GB
  • With parts like:
    • 11.5GB
    • 7.5GB
    • 320MB

---

Technical explanation: why multiple models are needed

The “under the hood” pipeline is broken into three major components:

1. Text encoder (text encoding / “text decryption” concept)

   • Converts prompt text into numeric embeddings the image model understands.

2. Image/generation model (the main diffusion/latent generator)

   • Uses the embeddings to generate an image representation in latent space.

3. VAE / decoder (“V”)

   • Converts from latent space into visible pixels.

---

Running workflows and understanding progress

• Press the Run button to execute.
• During execution:
  • Overall progress appears in the top-right
  • Per-node progress bars appear on the nodes themselves
• The first generation after installing/loading models is slower because models must load into memory.
• Subsequent runs are faster.
• Tip shown: replace Save Image with Preview Image if you don’t want automatic saving.

---

Prompting and model strengths/weaknesses

• Different components/models can handle different languages or tasks differently.
• Example behavior:
  • A text encoder model may understand Russian,
  • but another model component (the z-image generator) might generate poor Russian text.
• This is why some workflows require different models for different roles.

---

Multi-image editing workflow (Load Image + seed/CFG behavior)

Template-based demo

• The next demo uses a more capable model/workflow from Image templates.

Missing-input errors

• “Load image” nodes remain red until images are uploaded.

Editing approach

• Deactivate nodes with Ctrl+B.
• Uses multiple images for guidance (example described: original ship render + reference style image).

Seed / regeneration behavior

• A key generation rule is highlighted:
  • If the seed/noise value (“conditionally random number”) doesn’t change, parts of the workflow may not regenerate.
• Result: only certain sections may update until you change the right seed/conditions.

---

Upshot: converting a simple 3D render into a cinematic image

• The example begins with a “Dune-like ship in ice” scenario.
• Quality improves by:
  • using additional references
  • tweaking prompts
  • rerunning once the prompt/image setup is correct
• Larger models generally make generation slower.

---

Optional automation: using LLM-style templates to rewrite prompts

• You can load text/LLM prompt models via templates (like “lm”).
• Workflow flow:
  • generate a structured prompt using earlier output
  • paste that structured prompt into the image prompt
• Prompting emphasis:
  • request a structured prompt
  • avoid introductory phrases so the model doesn’t output extra text

---

Saving workflows as reusable projects

• ComfyUI supports saving:
  • generated images
  • the entire workflow
• Main advantage described:
  • If you move an output image into a new (empty) workflow page, ComfyUI can recreate the full process, including conditions and even seed, producing consistent results when re-run.

---

Main speakers / sources

• Main speaker: The presenter (name not provided in the summary)
• Software source: ComfyUI and conf.org (installation source mentioned)
• Model aggregation references: Various aggregator names listed at the start (e.g., Kre, Hixfield, Freepic, Vivi, Synx, Flashboards)