{"course":{"slug":"ai-images","title":"Image Generation, in Practice","url":"https://addaly.com/learn/ai-images","level":"beginner","track":"create"},"lesson":{"slug":"cost-and-batching","title":"You are paying to make decisions you could make cheaply","url":"https://addaly.com/learn/ai-images/cost-and-batching","position":11,"of":69,"minutes":8,"answer":"Choose the frame on cheap low-step renders and spend full quality only on the winner; after about six full-quality attempts the remaining distance is editing work, not generation.","keyIdea":"Choose the frame on cheap low-step renders and spend full quality only on the winner; after about six full-quality attempts the remaining distance is editing work, not generation."},"body":{"format":"markdown","text":"## Where the money actually goes\n\nA freelancer buys a month of credits, spends two days on one hero image, runs out, and has produced nothing usable. This is the standard failure and it is not about the price of the plan. It is about rendering at final quality while still deciding what the image is.\n\nAlmost every compositional decision (where the subject sits, which of four poses reads best, whether the horizon is high or low, which of three colour moods works) is visible at 512 pixels and 12 steps. That render costs a fraction of a full-quality one. Making those decisions at full quality is like printing every draft of a document on photo paper.\n\n## The ladder\n\n1. **Explore small and cheap.** Batches of four to eight at low resolution and low step count. Look for a composition, not for detail. Note the seed of anything promising.\n2. **Re-render the winner properly.** Same seed, same prompt, full resolution and step count. The composition transfers. The texture and detail arrive now.\n3. **Inpaint the defects**, one region at a time, at full resolution.\n4. **Upscale once**, at the very end.\n\nNine out of ten renders happen at stage one, and stage one is where the cost per render is lowest. That is the whole trick.\n\n## Step count is where people burn money out of superstition\n\nSampler steps cost linearly and stop buying you anything fairly early. With modern samplers, most of the image is settled by 20-30 steps; beyond that you are paying for changes you cannot see. Some distilled models are designed for four to eight steps and get worse if you push them higher.\n\nPeople run 80 steps because a tutorial in 2022 said so, about a different sampler on a different model. Run the same seed at 15, 25, 40 and 80 once, for your model, and then stop guessing forever.\n\nGuidance scale has the same shape. Too low and the image ignores the prompt; too high and colours blow out and the image goes crunchy. The sweet range is narrower than the slider suggests.\n\n## Batch of four, not one at a time\n\nGenerate four variants at once rather than one, four times. Two reasons. You judge better by comparison than in isolation: a single image looks fine until you see a better one next to it. And on most hosted tools a batch is cheaper per image than separate requests, because the setup cost is shared.\n\n## Keep a log, or you will lose the good one\n\nFor anything you might return to: prompt, negative prompt, model and version, seed, steps, guidance, sampler, and the file. A plain text file next to the images is enough.\n\nA prompt on its own is not reproducible. The same words on a different model, or the same model at a different guidance value, give you a different picture. People post prompts online as if they were recipes; without the rest of the settings they are a list of ingredients with no quantities.\n\n## Local versus hosted, roughly\n\nHosted generation sits in the region of a few cents an image on most platforms, so a thousand images is tens of dollars. Running locally is your electricity plus a graphics card you may not own, and a used card capable of comfortable SDXL work has cost a few hundred dollars for a while now.\n\nBreak-even is somewhere in the low thousands of images, so for most people hosted is the right answer financially, and local is the right answer for control and privacy rather than for cost.\n\nIf you have no money at all, the free path is complete: free daily allowances on several hosted tools, free GPU hours on Colab and Kaggle, and public Spaces on Hugging Face. Do your exploration on the free daily allowance, and spend paid credits only on final renders. That habit alone multiplies what a small budget buys.\n\nAnd on a preset platform, compare model prices before committing a batch, as lesson two showed, the cheap model and the premium model within one platform can differ by seven or eight times for output nobody can distinguish at the size it will be used.\n\n## Know when to stop generating\n\nIf six full-quality attempts at the same thing have not produced it, a seventh will not either. The model has shown you its distribution. The remaining distance is editing work (inpainting, compositing, a tone pass) and doing that is faster and cheaper than continuing to roll.\n\nSet a number before you start. When you hit it, move to the editor.\n\nToday: find the steps setting in whatever you use, run one seed at four different step counts, and see where the improvement stops for your model. Most people discover they have been paying double."},"check":{"question":"A freelancer exhausts a month's credit allowance in two days on a single hero image and still does not have a usable frame. What is the most useful correction?","options":[{"text":"The plan is too small for professional work and they should move to the next tier up.","correct":false,"tempting":"Buys more of a wasteful process instead of changing the process."},{"text":"The composition should have been chosen from batches of cheap, small, low-step renders before anything was rendered at final quality, and after roughly six full-quality attempts the remaining distance is editing work rather than generation.","correct":true},{"text":"They should move to running an open model locally, where images cost nothing per generation.","correct":false,"tempting":"Shifts the cost onto hardware without reducing the number of wasted full-quality renders."},{"text":"A higher step count would have reached an acceptable image in fewer attempts.","correct":false,"tempting":"Assumes quality keeps scaling with steps well past the point where the sampler has converged."}]},"prev":{"slug":"a-reproducible-setup","title":"The setup that still works in six months","url":"https://addaly.com/learn/ai-images/a-reproducible-setup"},"next":{"slug":"what-the-image-is-for","title":"Start at the placement, not the picture","url":"https://addaly.com/learn/ai-images/what-the-image-is-for"},"authorship":"Founder and series editor: Dr Mustafa Mun. Written by AI models to a written standard. Exam keys checked blind by a second model: 3,665 of 3,666 agreed. Academic review invited.","review":{"status":"Draft, open for academic review"},"cite":"Addaly (2026) 'You are paying to make decisions you could make cheaply', in Image Generation, in Practice. Edited by M. Mun. Addaly. Available at: https://addaly.com/learn/ai-images/cost-and-batching (Accessed: 31 October 2026).","licence":{"name":"CC BY-NC-SA 4.0","url":"https://creativecommons.org/licenses/by-nc-sa/4.0/"},"reader":{"schema":1,"title":"You are paying to make decisions you could make cheaply","courseTitle":"Image Generation, in Practice","module":{"index":1,"slug":"the-kit","title":"The kit, and what each piece of it costs you"},"answer":null,"keyIdea":"Choose the frame on cheap low-step renders and spend full quality only on the winner; after about six full-quality attempts the remaining distance is editing work, not generation.","hero":{"kind":"plate","image":null,"alt":null,"caption":null},"blocks":[{"t":"h2","id":"where-the-money-actually-goes","runs":[{"text":"Where the money actually goes"}]},{"t":"p","runs":[{"text":"A freelancer buys a month of credits, spends two days on one hero image, runs out, and has produced nothing usable. This is the standard failure and it is not about the price of the plan. It is about rendering at final quality while still deciding what the image is."}]},{"t":"p","runs":[{"text":"Almost every compositional decision (where the subject sits, which of four poses reads best, whether the horizon is high or low, which of three colour moods works) is visible at 512 pixels and 12 steps. That render costs a fraction of a full-quality one. Making those decisions at full quality is like printing every draft of a document on photo paper."}]},{"t":"h2","id":"the-ladder","runs":[{"text":"The ladder"}]},{"t":"ol","start":1,"items":[[{"text":"Explore small and cheap.","b":true},{"text":" Batches of four to eight at low resolution and low step count. Look for a composition, not for detail. Note the seed of anything promising."}],[{"text":"Re-render the winner properly.","b":true},{"text":" Same seed, same prompt, full resolution and step count. The composition transfers. The texture and detail arrive now."}],[{"text":"Inpaint the defects","b":true},{"text":", one region at a time, at full resolution."}],[{"text":"Upscale once","b":true},{"text":", at the very end."}]]},{"t":"p","runs":[{"text":"Nine out of ten renders happen at stage one, and stage one is where the cost per render is lowest. That is the whole trick."}]},{"t":"h2","id":"step-count-is-where-people-burn-money-out-of-superstition","runs":[{"text":"Step count is where people burn money out of superstition"}]},{"t":"p","runs":[{"text":"Sampler steps cost linearly and stop buying you anything fairly early. With modern samplers, most of the image is settled by 20-30 steps; beyond that you are paying for changes you cannot see. Some distilled models are designed for four to eight steps and get worse if you push them higher."}]},{"t":"p","runs":[{"text":"People run 80 steps because a tutorial in 2022 said so, about a different sampler on a different model. Run the same seed at 15, 25, 40 and 80 once, for your model, and then stop guessing forever."}]},{"t":"p","runs":[{"text":"Guidance scale has the same shape. Too low and the image ignores the prompt; too high and colours blow out and the image goes crunchy. The sweet range is narrower than the slider suggests."}]},{"t":"h2","id":"batch-of-four-not-one-at-a-time","runs":[{"text":"Batch of four, not one at a time"}]},{"t":"p","runs":[{"text":"Generate four variants at once rather than one, four times. Two reasons. You judge better by comparison than in isolation: a single image looks fine until you see a better one next to it. And on most hosted tools a batch is cheaper per image than separate requests, because the setup cost is shared."}]},{"t":"h2","id":"keep-a-log-or-you-will-lose-the-good-one","runs":[{"text":"Keep a log, or you will lose the good one"}]},{"t":"p","runs":[{"text":"For anything you might return to: prompt, negative prompt, model and version, seed, steps, guidance, sampler, and the file. A plain text file next to the images is enough."}]},{"t":"p","runs":[{"text":"A prompt on its own is not reproducible. The same words on a different model, or the same model at a different guidance value, give you a different picture. People post prompts online as if they were recipes; without the rest of the settings they are a list of ingredients with no quantities."}]},{"t":"h2","id":"local-versus-hosted-roughly","runs":[{"text":"Local versus hosted, roughly"}]},{"t":"p","runs":[{"text":"Hosted generation sits in the region of a few cents an image on most platforms, so a thousand images is tens of dollars. Running locally is your electricity plus a graphics card you may not own, and a used card capable of comfortable SDXL work has cost a few hundred dollars for a while now."}]},{"t":"p","runs":[{"text":"Break-even is somewhere in the low thousands of images, so for most people hosted is the right answer financially, and local is the right answer for control and privacy rather than for cost."}]},{"t":"p","runs":[{"text":"If you have no money at all, the free path is complete: free daily allowances on several hosted tools, free GPU hours on Colab and Kaggle, and public Spaces on Hugging Face. Do your exploration on the free daily allowance, and spend paid credits only on final renders. That habit alone multiplies what a small budget buys."}]},{"t":"p","runs":[{"text":"And on a preset platform, compare model prices before committing a batch, as lesson two showed, the cheap model and the premium model within one platform can differ by seven or eight times for output nobody can distinguish at the size it will be used."}]},{"t":"h2","id":"know-when-to-stop-generating","runs":[{"text":"Know when to stop generating"}]},{"t":"p","runs":[{"text":"If six full-quality attempts at the same thing have not produced it, a seventh will not either. The model has shown you its distribution. The remaining distance is editing work (inpainting, compositing, a tone pass) and doing that is faster and cheaper than continuing to roll."}]},{"t":"p","runs":[{"text":"Set a number before you start. When you hit it, move to the editor."}]},{"t":"p","runs":[{"text":"Today: find the steps setting in whatever you use, run one seed at four different step counts, and see where the improvement stops for your model. Most people discover they have been paying double."}]}],"check":{"question":[{"text":"A freelancer exhausts a month’s credit allowance in two days on a single hero image and still does not have a usable frame. What is the most useful correction?"}],"options":[{"text":[{"text":"The plan is too small for professional work and they should move to the next tier up."}],"correct":false,"tempting":[{"text":"Buys more of a wasteful process instead of changing the process."}]},{"text":[{"text":"The composition should have been chosen from batches of cheap, small, low-step renders before anything was rendered at final quality, and after roughly six full-quality attempts the remaining distance is editing work rather than generation."}],"correct":true,"tempting":null},{"text":[{"text":"They should move to running an open model locally, where images cost nothing per generation."}],"correct":false,"tempting":[{"text":"Shifts the cost onto hardware without reducing the number of wasted full-quality renders."}]},{"text":[{"text":"A higher step count would have reached an acceptable image in fewer attempts."}],"correct":false,"tempting":[{"text":"Assumes quality keeps scaling with steps well past the point where the sampler has converged."}]}],"order":[0,2,1,3],"why":null},"prev":{"slug":"a-reproducible-setup","title":"The setup that still works in six months","minutes":7},"next":{"slug":"what-the-image-is-for","title":"Start at the placement, not the picture","minutes":8}}}