ComfyUI From Zero - Part 1 - homework answers: workflow 03

Workflow 03 is workflow 01 with three things broken on purpose: two stop it running, the third lets it run and gives
you a picture that looks wrong. Put it next to workflow-01-first-image.json and compare them node by node: those three
are the only differences. workflow-03-fixed.json is the repaired workflow.

Tried it yourself first? Good. Here they are.


BREAK 1 - the model name (stops the run)

What you see   Press Run and a red box says "Prompt outputs failed validation":
                   CheckpointLoaderSimple: Value not in list: ckpt_name: 'definitely_not_a_real_model.safetensors' not in [...]
               The Load Checkpoint node gets a red outline. (Newer ComfyUI versions already warn you about the missing
               model when you open the workflow.)
Why            A loader can only load a file that is in its folder, here ComfyUI/models/checkpoints/. This name
               belongs to no file at all. It's error 2 of the eight in Module 9, "Value not in list".
The fix        Click the ckpt_name box in Load Checkpoint and pick z_image_turbo_aio.safetensors.
               Don't have it yet? The download link is in the note of workflow 01 and in READ ME.txt. Copy the file
               in, press R, then pick it.


BREAK 2 - the missing cable (stops the run)

What you see   Next to (or after) break 1, the error says:
                   KSampler: Required input is missing: latent_image
               The KSampler gets a red outline. Its latent_image input has no cable, and nothing leaves the LATENT
               output of Empty Latent Image.
Why            The sampler needs a latent to start from, the blank photo paper. An input on the left side of a node
               must be connected before the workflow can run. It's error 8 in Module 9, "Required input is missing".
The fix        Drag a cable from the LATENT output of Empty Latent Image to the latent_image input of the KSampler.


BREAK 3 - denoise 0.25 (runs, but the picture is wrong)

What you see   No error. The run finishes and saves an image, but it isn't the red camera on the desk: you get a
               flat, washed-out, blurry picture, at most a faint ghost of the camera.
Why            Denoise is how much of the starting latent the sampler is allowed to change, from 0 to 1. Below 1.0
               it keeps part of what it starts from, which is what you want for image-to-image. An empty latent has
               nothing worth keeping, so at 0.25 the sampler only does the last quarter of the work and the picture
               never gets drawn. The settings look normal at a glance, which is why this one is the hard one.
The fix        Set denoise in the KSampler back to 1.00. Text-to-image from an Empty Latent Image always runs at
               denoise 1.0.


With all three fixed, workflow 03 is workflow 01 again: the same model, seed 42, steps 8, CFG 1, the same picture.
