{
  "$type": "site.standard.document",
  "bskyPostRef": {
    "cid": "bafyreicburpxle3yubktwxmcpjgzbn4zum66yvalzfwjn36mricgg6prtq",
    "uri": "at://did:plc:pgryn3ephfd2xgft23qokfzt/app.bsky.feed.post/3miy62ixa4yp2"
  },
  "path": "/t/dont-even-know-where-to-start/175065#post_5",
  "publishedAt": "2026-04-08T08:41:46.000Z",
  "site": "https://discuss.huggingface.co",
  "tags": [
    "try out the HF Spaces demo to see which model actually suits your needs",
    "ComfyUI",
    "Google Research",
    "GitHub",
    "YouTube"
  ],
  "textContent": "Yeah. Google Colab free tier is good one. I also use the Tesla T4 16GB on Colab Free on a daily basis for experimental purposes.\n\nWell, as long as you have 8GB of VRAM, there are models that work just fine as long as they aren’t the latest ones, so I think it’s safer to start with a lower-end model to get the hang of things, **decide whether you’ll use ComfyUI or not, and then choose your GPU**.\nAn 8GB GPU is plenty for practice. You might also want to try out the HF Spaces demo to see which model actually suits your needs.\n\nThe **Windows environment is particularly inconvenient when it comes to AI, and there are so many pitfalls to watch out for when choosing a GPU** … (I’m a Windows user myself. )\n\n* * *\n\nStart with **ComfyUI + SD 1.5** , not with FLUX or a giant multi-image workflow.\n\nThat is the safest fit for your actual setup: **8GB VRAM on Windows** plus **Colab Free as a backup**. ComfyUI’s official beginner docs start with built-in workflow templates, model installation, and a first working run. The official image-to-image tutorial uses **SD 1.5** (`v1-5-pruned-emaonly-fp16.safetensors`), and the official inpainting tutorial uses a dedicated **SD 1.5 inpainting** checkpoint (`512-inpainting-ema.safetensors`) and explicitly says it gives more natural inpaint results than a normal SD 1.5 checkpoint. (ComfyUI)\n\n## What is realistic for your hardware\n\n### Best first local option\n\n**SD 1.5 in ComfyUI** is the best place to begin. It is directly used in the official beginner image-to-image and inpainting guides, which means you get a supported path and lighter workflows at the same time. (ComfyUI)\n\n### Second step\n\n**SDXL** is plausible on modest NVIDIA hardware, but I would not make it lesson one in ComfyUI on exact 8GB. It is better as a step after you already understand image-to-image and inpainting. This is an inference from your 8GB limit plus the fact that the official beginner docs use SD 1.5 for the basic edit tutorials rather than SDXL. (ComfyUI)\n\n### Later experiment\n\n**FLUX.2 Klein 4B Distilled** is interesting, but not the safest first local target on exact 8GB. ComfyUI’s own guide lists the distilled 4B model at about **8.4GB VRAM** and the base 4B model at about **9.2GB VRAM** on an RTX 5090. The same guide says Klein supports **style transforms** , **semantic edits** , **object replacement/removal** , **multi-reference composition** , and **iterative edits**. That makes it a good later model for your goals, but a tight fit for local 8GB. (ComfyUI)\n\n### Colab Free\n\nUse **Colab Free as a backup lane** , not as your main learning lane. Google’s FAQ says free Colab has **dynamic usage limits** , **no guaranteed or unlimited resources** , varying GPU types, idle timeouts, and notebooks that can run for **at most 12 hours** depending on usage and availability. (Google Research)\n\n## My blunt recommendation\n\nFor your setup, I would rank the starting paths like this:\n\n  1. **ComfyUI + SD 1.5 image-to-image + SD 1.5 inpainting**\n  2. **ComfyUI + SDXL** , after the basics feel normal\n  3. **FLUX.2 Klein 4B Distilled** , preferably later or on a favorable Colab session (ComfyUI)\n\n\n\n## ComfyUI or Forge Classic\n\nFor **learning** , I would still start with **ComfyUI**.\n\nWhy:\n\n  * ComfyUI’s official docs are now structured around **Templates** , **first generation** , **image-to-image** , and **inpainting** , so there is a clean beginner path. (ComfyUI)\n  * ComfyUI is explicitly a **graph / nodes / flowchart** interface, which is exactly why it scales better once you move from “edit one image” to “build one image from two references.” (GitHub)\n  * The official examples repo says all example images contain **workflow metadata** , so you can drag them into ComfyUI and recover the workflow used to make them. That is extremely beginner-friendly once you know the first few nodes. (GitHub)\n\n\n\n**Forge Classic** is still a valid alternative if you want a more familiar WebUI-style interface. Its README says it is built on top of the original AUTOMATIC1111-style WebUI, focuses on optimization and usability, and currently supports **Flux.2-Klein 4B/9B** but **not FLUX.2 Dev**. The same README also notes that Klein there does **not support regular img2img** and “will always edit.” (GitHub)\n\nSo my practical split is:\n\n  * **Use ComfyUI** if you want the best long-term path for editing + two-image workflows. (GitHub)\n  * **Use Forge Classic** if you mainly want a simpler A1111-style UI and are okay staying in a more traditional WebUI workflow. (GitHub)\n\n\n\n## Step by step: the safest path\n\n### Step 1. Open Templates\n\nIn ComfyUI, open **Workflow Templates** from the sidebar or from **Workflow → Browse Workflow Templates**. The Templates browser is where ComfyUI puts its natively supported model workflows and example workflows. When you load a template, ComfyUI checks for missing models and prompts you to download them. (ComfyUI)\n\n### Step 2. Do one simple first run\n\nUse the official **Getting Started with AI Image Generation** guide. It is specifically written to cover workflow loading, model installation, and a first working image. It also explains that the default workflow usually loads automatically, and it shows how to load workflows from Templates or from images with workflow metadata. (ComfyUI)\n\n### Step 3. Learn image-to-image\n\nThis should be your first real task. The official image-to-image guide says it is used for style transfer, line art to realism, restoration, and colorizing old photos. It also says the workflow is very similar to text-to-image, just with an added reference image, which is exactly why it is a good beginner bridge. The key setting is `denoise`: lower values keep the result closer to the source image, higher values change it more. (ComfyUI)\n\n### Step 4. Learn inpainting\n\nAfter image-to-image, do inpainting. The official inpainting guide is about changing only the masked area, and it covers the **Mask Editor** and the **`VAE Encoder (for Inpainting)`** node. It also explicitly shows that the dedicated inpainting checkpoint gives better transitions than a normal SD 1.5 checkpoint. (ComfyUI)\n\n### Step 5. Only then try two-image work\n\nFor “make a new image from two images,” the most important shift is conceptual:\n\n  * image A can provide **subject/content**\n  * image B can provide **style/look**\n  * or both images can act as **references** for a new composition\n\n\n\nKlein is relevant here because the official guide says it supports **multi-reference composition** , but on your hardware I would treat that as a later step, not the first one. (ComfyUI)\n\n## The simplest model plan for you\n\n### Local Windows 8GB\n\nUse:\n\n  * **SD 1.5** for image-to-image\n  * **SD 1.5 inpainting model** for local masked edits\n\n\n\nThat matches the official beginner tutorials directly. (ComfyUI)\n\n### Colab Free\n\nUse Colab only when you want to test something heavier or more modern. Do not build your learning routine around it because the free tier is not predictable. (Google Research)\n\n### Klein\n\nTry **FLUX.2 Klein 4B Distilled** later, especially if you get a decent Colab session or upgrade local VRAM. It is a good model for your eventual goal set, but not the calmest way to start on exact 8GB. (ComfyUI)\n\n## Good beginner guides\n\nThese are the ones I would actually use:\n\n### Official first\n\n  * **Getting Started with AI Image Generation**. Best first run guide. (ComfyUI)\n  * **Workflow Templates**. Best place to find safe starter workflows. (ComfyUI)\n  * **Image-to-Image**. Best first editing guide. (ComfyUI)\n  * **Inpainting**. Best first local-edit guide. (ComfyUI)\n  * **ComfyUI Examples**. Best place to inspect working workflows because the images contain workflow metadata. (GitHub)\n\n\n\n### Video guides\n\n  * **Pixaroma – Learn ComfyUI From Scratch**. It is explicitly presented as a complete beginner course for learning ComfyUI from scratch. (YouTube)\n  * **Scott Detweiler – ComfyUI playlist**. Long-running playlist focused on local Stable Diffusion and ComfyUI workflows. (YouTube)\n\n\n\n## The biggest beginner mistakes to avoid\n\nDo not start from a random workflow JSON or a giant image with missing custom nodes. Start with **Templates** and official examples. That is exactly what ComfyUI’s getting-started docs and examples repo are set up for. (ComfyUI)\n\nDo not install a pile of custom nodes before you have one working image-to-image flow and one working inpaint flow. The official beginner path is core workflows first, models next, then more advanced paths. (ComfyUI)\n\nDo not make Klein your first local test on exact 8GB just because it is modern. Its own published VRAM figure is already sitting right at the edge of your card class. (ComfyUI)\n\n## The short version\n\nFor **your exact setup** , start like this:\n\n  1. **ComfyUI**\n  2. **Official Templates**\n  3. **SD 1.5 image-to-image**\n  4. **SD 1.5 inpainting**\n  5. **Only then** test SDXL\n  6. **Only then** test FLUX.2 Klein 4B Distilled, preferably on a good Colab session or stronger VRAM (ComfyUI)\n\n\n\nThat is the least frustrating path from “I installed this thing” to “I can edit images and understand what the workflow is doing.”",
  "title": "Dont even know where to start!"
}