Week 2: The Loop

The basic cycle every generator runs on — the interface against the workflow underneath it, rolling and selecting rather than prompting once, and the seed as the address of a specific result.

1 March 2027

Every generator this course touches runs the same loop underneath whatever interface sits on top of it: set some parameters, generate, look at what came back, decide whether to keep it, and either stop or go again. The interface is the part you see — a form, a chat box, a graph of nodes; the workflow is the actual sequence of operations underneath it, and the two are not the same thing. For the first six weeks this course supplies the workflow, as a starting graph, so the loop itself is the whole lesson.

What one roll is specified by, on the sampler the supplied graph runs.
seed0 to 0xffffffffffffffff — the random seed used for creating the noise. Change it and a different roll has been generated; hold it, and one other parameter is being tested on its own.
stepsHow many denoising steps the sampler runs before it hands a result back. Default 20.
cfgHow hard the sampler is pushed toward the prompt rather than toward whatever the model would produce unprompted. Default 8.0, with 6 to 8 recommended.
samplereuler, euler_ancestral, heun, dpmpp_2m, dpmpp_2m_sde or dpmpp_3m_sde, each paired with a scheduler: simple, sgm_uniform, karras, exponential, ddim_uniform, beta or normal. The list is read out of the running software and moves; this one was read in September 2026.

Source: ComfyUI, KSampler

Rolling is not re-prompting

Running the same prompt more than once is called rolling, and it is not a sign the first attempt failed — it’s the default way this loop is used. A generator is a sampler, not an oracle: the same prompt with a different random draw produces a genuinely different result, and the skill this course credits is choosing well among several rolls, not writing a prompt so precise it only needs one. Selection is the actual job. Generation is just what makes selection possible.

The seed as an address

A generation seed is not a style knob; it’s an address. On the same model, sampler, and software version, the same seed with the same parameters returns to the same result, which means a seed plus its full parameter list is the only thing that actually specifies a roll. Change one parameter and keep the seed, and a specific variable is being tested in isolation; change the seed and a different roll has been generated regardless of what else stayed the same. Log both, every time, or the roll that looked right yesterday is unfindable tomorrow.

  1. Seed 66740, served at /studio/week02-t1.mp4 and /studio/week06-t4.mp4.

  2. Seed 66741, served at /studio/week02-t2.mp4 and /studio/week07-t4.mp4.

Two rungs of two later ladders are not like these rolls; they are these rolls. Same seed, same first frame, same prompt file, so the same take — and the library keeps one recording and answers at both addresses. Switching changes the address, not the picture.

The production log

The log is not paperwork attached to the work; it is the record that turns a good roll into a repeatable one. A usable log names the workflow graph in use, the seed, and every parameter that could plausibly change the output — enough that someone who wasn’t in the room could load the same graph and land on a comparable result. A roll without a log is a lucky accident the moment the tab closes.

instrument  The Loop
model       MiniMax H3
mode        I2VA
resolution  576x1024
fps         24
prompt      inputs/week02/prompt.md

tier  seed   output
t1    66740  week02-t1.mp4
t2    66741  week02-t2.mp4
t3    66742  week02-t3.mp4
t4    66743  week02-t4.mp4
t5    66744  week02-t5.mp4
The rolls in this week's ladder as the log records them: one prompt file, one set of settings, and the seed that addresses each roll.

The ladder

This week’s ladder runs sideways rather than up, because the skill being graded is selection, not escalating difficulty. The exercise is five rolls of exactly one prompt, on the graph the course supplies, with nothing changed between them but the random draw. Each of the five gets its own checklist verdict against the same four checks — hands, physics, text in frame, and overall framing — logged as pass or fail rather than a general impression. Climbing this ladder means getting more of the five to clear the checklist, and being able to say precisely which check failed on the ones that didn’t.

seed 66740

One prompt, five seeds: watch how much of the shot the seed alone decides.

seed 66741

Compare the rain direction and the hair drift against t1 - same words, different roll.

seed 66742

Watch whether the coat collar and the amber lights stay where the still put them.

seed 66743

Look for the seeds that drift the face: identity is the first thing a bad roll loses.

seed 66744

Pick the take you would keep; that choice is the whole loop.

Recorded on the convener's rig with MiniMax H3 (local weights), keyframes with NoobAI-XL, reference stills for The Ninth Head with ChatGPT's image generator. Nothing is generated in the browser. Run it in the Studio.

Before class

Install the workflow engine if it isn’t already running from last week, and load the graph this course supplies for the semester’s opening weeks.

This week’s exercise

Roll one prompt five times on the supplied graph, log the seed and parameters for every roll, and write a pass or fail verdict for each against the hands, physics, text, and framing checklist. Bring all five rolls, their logs, and your five verdicts to Wednesday’s Dailies.

Reading

Teaching team