# Lab C — Denoise in public **Time:** 45–60 minutes · **Cost:** $0 local, ~$1 rented **Needs:** 6GB VRAM, or Apple Silicon with 16GB unified, or a rented GPU **Makes visible:** generation is a walk from noise, not a lookup. Chapter 5. --- ## Quickest way to run this Double-click **`RUN-THIS-Windows.bat`** (Windows) or **`RUN-THIS-Mac-Linux.command`** (Mac) in this folder. It sets everything up the first time and then runs the lab. Nothing else needed. *Mac: if it refuses to open, right-click it and choose Open, then Open again. Only needed once.* Prefer to type the commands yourself? Carry on below. --- ## Setup **Run these one line at a time.** Do not paste the fence lines (the triple backticks) — they are markdown, not commands. Pasting a whole block at once is the single most common way this lab fails before it starts. ### Windows (PowerShell) ```powershell py -m venv .venv .\.venv\Scripts\python.exe -m pip install -r requirements.txt ``` ### macOS / Linux ```bash python3 -m venv .venv ./.venv/bin/python -m pip install -r requirements.txt ``` ### Verify before you continue This must print a path ending in `.venv`: ```powershell .\.venv\Scripts\python.exe -c "import sys; print(sys.prefix)" ``` ```bash ./.venv/bin/python -c "import sys; print(sys.prefix)" ``` If it prints a conda path, a system Python, or anything else, **stop**. Your packages installed into the wrong environment and nothing below will work. ### Running the lab Every `python` in this README means the venv's interpreter: ```powershell .\.venv\Scripts\python.exe denoise_in_public.py ``` ```bash ./.venv/bin/python denoise_in_public.py ``` Calling the interpreter by path is deliberate. It needs no `activate`, so PowerShell's execution policy cannot block it, and there is no way to silently install into the wrong environment — which is what `source .venv/bin/activate` does on Windows, where it fails without stopping the script that follows. First run downloads ~4GB of model weights. Once only. **No GPU?** See `../SETUP.md` for renting one for an hour (about $0.35–0.69). An RTX 4090 or an A5000 is plenty. Do not rent an H100 for this. **CPU-only** works but is genuinely painful — several minutes per image. If you're going that route, use `--steps 8` and accept ugly results; you're looking at the *sequence*, not the quality. --- ## Why an old model The default is Stable Diffusion 1.5, from 2022. There are far better generators now and we are not using them on purpose: - One text encoder instead of two, which makes the Lab D interpolation about fifteen lines instead of fifty. - Runs on 4GB of VRAM. - Every mechanism in Chapter 5 is present and visible. This is the teaching model, not the quality model. Newer families want 10–12GB minimum and complicate the code without changing a single thing you're meant to learn. Go look at current models afterwards, separately. --- ## Part one — the walk ```bash python denoise_in_public.py ``` Generates one image and saves a PNG at **every** denoising step into `output/trace/`. Now open that folder and scan the files in order. This is the lab. It is not complicated and it doesn't need to be. **The thing nobody tells you in advance: notice how early the composition is fixed.** Within the first fifth of the steps — before there's anything you'd call detail — the overall structure is already committed. Everything after is refinement inside a decision that has already been made. That has a direct product consequence. If a generation is going to come out wrong in composition, it was wrong within the first few steps, and adding steps at the end will not save it. ## Part two — three sweeps **Step count.** Where does it stop improving? ```bash python denoise_in_public.py --sweep-steps ``` Same seed, same prompt, at 2/5/10/20/35/50 steps. Line them up. Somewhere between 20 and 35 the returns stop. That point is your quality/cost frontier, measured rather than guessed — and it is the same *shape* of curve you'll meet again in Lab F with training examples on the x-axis instead of steps. **Guidance.** What CFG scale actually does. ```bash python denoise_in_public.py --sweep-guidance ``` At 1.0 the model wanders wherever it likes and barely registers your words. At 7.5 you get adherence with natural results. At 25 it's contorted and oversaturated — you amplified the text direction so hard that you pushed the latent outside the region where the decoder produces anything sane. You are watching a vector subtraction get multiplied. That slider is in every image UI on earth and almost nobody using it knows what it's doing. **Seeds.** Same words, different starting point. ```bash python denoise_in_public.py --sweep-seeds ``` Every one of those images got identical instructions. All the variation comes from where the walk started. Keep that distinction: **"the prompt was ambiguous" and "the walk started somewhere else" are different diagnoses with different fixes**, and teams confuse them constantly. If seeds vary wildly on a prompt you thought was specific, that tells you the prompt is underconstraining the region — which is a prompt problem. If they're all similar and all wrong, that's a model problem. --- ## Troubleshooting **CUDA out of memory** — add `--steps 15`, or use `stabilityai/sdxl-turbo` with `--steps 4`, or rent a GPU. The script already enables attention slicing on CUDA. **`safety_checker` warnings** — expected and harmless; it's disabled deliberately so it can't silently blank an image mid-sweep and confuse your comparison. **Model 404s** — Hugging Face repo IDs get renamed. `../CURRENT.md` lists the current ID and a mirror. **macOS: very slow or black images** — MPS support varies by torch version. Try `--device cpu` to confirm the pipeline works at all, then upgrade torch. **Images are all noise** — you're looking at early steps. Check `final.png`.