Image Generation, in Practice

Farhan and Divya run a studio out of a converted flat. Between them they had delivered two hundred and six images for the co-operative over eleven weeks: packaging mock-ups, market-stall backgrounds, a series of illustrated farmer portraits for the rebranded milk cartons, and a set of forty social cards. All of it had been generated locally, on a second-hand card Farhan had bought precisely so that per-image cost would stop being a line in their quotes.

The model was FLUX.1 [dev]. They had chosen it in week one because it held type better than anything else they could run, the samples on the model page were the best they had seen, and the download was free and took twenty minutes. Nobody read the licence file that came down with the weights. It was sitting in the folder the whole time.

In week eleven the co-operative’s new marketing head asked a reasonable question: for the packaging artwork, which is going to a printer and onto half a million cartons, what are we licensed for? Divya opened the model card to copy a line into the email and found that FLUX.1 [dev] is released under a non-commercial licence. The weights may be used for research and personal work. Generating output that a client pays for is not that.

Two things were immediately clear and one was not. The first: this was a breach of the model licence, and it was entirely separate from any question about who owns the pictures. The second: it was not the client’s fault and the client had no exposure they knew about. The third, the unclear one, was what to do.

The options were not equal.

Say nothing. Nobody had noticed in eleven weeks. The licence is not policed by anybody who would plausibly look at a dairy co-operative in Madhya Pradesh. The studio’s reputation, their retainer and their next quarter all survive intact. The cost is that Divya would have to answer the marketing head’s direct question with something that was not true, in writing, and then live with that sentence existing in an inbox for as long as the account did.

Rebuild everything. Re-generate two hundred and six images on Apache-licensed weights (FLUX.1 [schnell] or Qwen-Image, both of which permit commercial use), then repair, composite and grade them all again. Farhan estimated it at nine days of work they would not be paid for, and it would push two deliverables past dates the co-operative had already booked print slots against.

Rebuild only what is exposed. The packaging artwork and anything going to print is the part with real exposure and volume behind it. The social cards are ephemeral. That is about thirty images rather than two hundred, and it is defensible if you can say where the line was drawn and why.

Buy the licence. Black Forest Labs sells a commercial licence for the dev weights. The studio had never priced it because they had never thought they needed to.

Divya’s position was that the third option was a rationalisation dressed as a plan: the licence does not distinguish between a carton and a social card, and drawing a line by commercial risk is deciding which breaches matter. Farhan’s position was that nine unpaid days would take the studio’s cash below the point where they could pay the rent in April, and that a rule you cannot afford to follow completely is still worth following partly.

The argument took a day. It was not really about the licence by the end of it. It was about whether a two-person studio that does not check terms before it starts is a studio a client should hire, and whether either of them wanted to keep working in a way that depended on nobody asking.

The brief arrived as one line in a WhatsApp message: something premium for admissions, like the big chains. The institute had two floors above a bank, one hundred and ninety students, and a genuinely good physics faculty. The director’s reference was a competitor’s brochure showing a landscaped campus, a library with a double-height ceiling, and eleven smiling students on a lawn.

Anusha, the designer, did the thing the brief lesson asks for before opening any generator: she asked where each image would appear, and she asked the fourth question: does anything real have to be in it.

The answers produced a list of eleven images. Four were straightforward: an abstract cover, two section dividers, a background for a results panel. Four were environments (a study scene, a library feel, a revision-at-night mood, a corridor) where nothing depicted a specific real place. And three were the problem: a photograph captioned our campus, a group described as our students, and a portrait row labelled our faculty.

The director wanted all eleven generated. His reasoning was not dishonest in his own mind. The building looks tired, he said, and students judge on photographs, and every institute in this town shows a campus that is nicer than it is. Photographing the actual premises would produce something that looked worse than the truth, because the corridor lighting is bad and the classrooms are full.

Anusha’s position was that the four environments were fine to generate, the four abstracts obviously so, and the three labelled ones were not. A generated building captioned our campus is not a stylistic choice; it is a claim about a place that does not exist, made to sixteen-year-olds and their parents deciding where to spend a year’s savings. The same for a room of generated students and a row of generated faculty.

The cost of saying so was real. The institute had been recommended to her, the fee was the largest single job in her quarter, and the director had made it clear that another designer would do it if she would not. She also knew that he was right about one thing: the competitor’s brochure almost certainly contained a generated lawn.

The cost of not saying so was also real, and it was not only ethical. Parents visit. A parent who has driven from Bundi on the strength of a photograph of a library, and then sees the actual first floor, is a parent who tells other parents. The image that was supposed to build trust is the one that removes it, and it does so at the worst possible moment, in person, with the director in the room.

So she wrote the spec before arguing. One page. Deliverables with pixel sizes. Native render ratios. The forbidden list: no legible brand names, no text other than type we set, no recognisable real landmarks. And then two lines that were the whole negotiation:

Photographed, not generated: the premises, the faculty, and any image captioned as depicting this institute.

Generated: environments, textures and abstract backgrounds that do not depict a specific place or person.

She took the spec to the meeting rather than an opinion, and she brought an alternative rather than a refusal: one afternoon with a phone, a window, and a white card, photographing four faculty members properly and the two best corners of the building at the right hour. Then the environments generated around them, which is a composite and not a fabrication.

The brief was agreed and unusually clear: one hero frame, 1200 by 1500, a line of houseboats moored at dawn, mist on the water, no people, room for a headline across the top third. The operator had asked for one specific thing: seven boats, because the fleet is seven boats and the number appears in the copy.

Ramya started properly. Slots, not a sentence. Medium, subject, action, setting, light, camera, finish. She fixed the seed once she had a composition she liked and ran the first test.

Six boats.

She added seven houseboats at the front of the prompt, where CLIP weights earlier tokens more heavily. Nine boats. She tried exactly seven wooden houseboats in a row. Five. She tried weighting the number. She tried spelling it as a numeral. She moved to a FLUX-family model on the theory that a T5 encoder handling longer natural language might bind the count. Eight. She went back, raised the guidance, lowered it, changed sampler, and generated another batch.

Forty minutes had gone. Every attempt was a competent, attractive picture of houseboats at dawn with the wrong number of them. She had two hours before she wanted to send a first look.

This is the moment the lesson calls the sunk-cost hour, and the diagnostic was sitting in front of her if she had been writing it down. Each failure was the same failure. Not a random scatter of problems (flat light in one, bad composition in another, a duplicated hull in a third) but the identical defect every time, at every setting, on two different architectures. A random failure means the model can produce what you want and you are sampling around it. A systematic one means the thing you want is outside what the model represents, and the twentieth attempt will look like the third.

Counting is a known systematic failure. There is no counting mechanism in a diffusion model. A number word shifts a distribution over how many things appear; below about four it usually lands and above it drifts.

Her options, once she named it, were three, and they had different costs.

Keep prompting. Cost: the remaining two hours, with no reason to expect a different outcome, and a first look that might not exist.

Change the brief. Frame the shot so only three boats are in view, receding into mist, with the rest implied. Cost: the number seven, which was the one thing the client had actually specified, and a conversation about why.

Change method. Generate the scene without a count at all, generate a single boat cleanly on its own, and composite the seven. Cost: perhaps ninety minutes of cutting out, scaling, matching perspective, painting reflections into the water, and matching grain, none of it glamorous, all of it predictable.

The third option also carried a second-order benefit she noticed only when she thought about it: a composited fleet is reusable. If the operator buys an eighth boat, that is a ten-minute change rather than a new hunt for a seed.

She also had one cheaper thing to do first, which she nearly skipped. She had never run a contact sheet on this brief. She had been judging one full-quality frame at a time for forty minutes, at ninety seconds a render, with commitment bias pulling her toward each one. Sixteen frames at twelve steps on her own checkpoint was ninety seconds in total.

The pitch needed four visuals of a branch interior that did not exist yet: the customer hall from the entrance, the same hall from the counter, a corner with the queuing system, and a manager’s cabin. The practice had the floor plan, the furniture schedule and two weeks.

Sameer, who does their visuals, started with prompts. The results were beautiful and useless. Each one was a plausible bank interior; none of them was this bank interior, and no two agreed about where the counter was or how high the ceiling ran. When he described the layout in words (the counter along the left, a waiting bank of six chairs facing it, a glazed cabin at the back right) the model bound about half of it and invented the rest.

He moved to sketching, which is the right instinct and the cheaper tool. Five minutes of biro, photographed on a phone, contrast raised in GIMP, fed as a scribble control at strength 0.7 ending at 0.65. The arrangement came back correct. The counter was on the left. The chairs were where he had drawn them.

And the room was subtly impossible. His drawn walls did not converge on a consistent vanishing point, and the conditioning had been faithful to his mistake. The senior partner, who has looked at interiors for twenty years, said it felt wrong within four seconds of seeing it and could not say why for another thirty. That is the specific failure of a sketch: it fixes arrangement and it does not fix perspective, and clients notice the difference precisely when the image is large on a screen in a meeting.

So the decision, at the start of week two.

Redraw the sketches more carefully. Sameer can construct perspective by hand; he learned it at college. Cost: two hours per view, four views, and a real chance of producing a different perspective error rather than none, because hand-constructed perspective is exactly as accurate as the person doing it at eleven at night.

Keep the sketch route and fix the geometry in repair. Cost: perspective is baked into every part of a frame at once. It is the category the keeper checklist puts under cannot be fixed by repair, and an hour spent discovering that is an hour gone.

Learn enough Blender to build a blockout. Not a model: grey boxes. Planes for floor and walls, cubes for the counter and the chairs, a cylinder for the queue post, a capsule where a person stands. Set a camera with a real focal length at a real eye height, render a depth pass, feed it as a depth control at 0.6 ending at 0.7, and prompt the materials and the light. Cost: an hour of a tutorial and an afternoon of fumbling, on a pitch with a deadline, by a person who had avoided Blender for six years.

The practice had one more consideration that decided it. The bank’s committee would ask for changes. They always do, and the changes are always move the counter and can we see more of the waiting area. On the sketch route, each of those is a new drawing and a new hunt for a seed. On a blockout, moving the counter is thirty seconds and a re-render, and the four views stay geometrically consistent with each other because they are four cameras in one scene.

The hero image had been approved on Tuesday. It was 4:5, generated at 896 by 1152 and finished properly: the product photographed against a window and composited into a generated workbench scene, contact shadow painted in, grain matched, forty minutes of repair on the bench edges and a stray reflection. The client had signed it off in writing.

At four o’clock on Thursday, two hours before the page went live, the client’s developer said the desktop hero slot was 16:9 and always had been. Nobody had asked the question at the start, and the spec, which Kiran had written, said Hero: 1200 × 1500 and nothing else.

Cropping a 4:5 frame to 16:9 removes the top and bottom, which is where the product and the light source were. The composition dies. So the real options were two, and both had a clock on them.

Outpaint the sides in steps. Enlarge the canvas by 256 pixels, mask the new area plus about 100 pixels of the existing image, inpaint at high denoise with a prompt describing what continues there, repeat until the ratio is right. That is roughly four steps each side. Each step risks perspective drift on the workbench edge, light drift as the new area gradually brightens, and the characteristic second-subject failure if the prompt still describes the product.

Regenerate at 1344 × 768 using the approved frame as an img2img reference at low strength, then re-composite the photographed product, repaint the contact shadow, and redo the grade. A clean 16:9 composition, and about ninety minutes of work that has already been done once.

Kiran chose outpainting, for a reason that was correct at four o’clock: the approved frame was approved, and a regeneration is a new picture that the client has not seen and might not like at six. Signed-off work is worth protecting when the clock is the binding constraint.

The execution nearly went wrong in exactly the documented way. On the first attempt he extended 400 pixels in one leap with the original scene prompt still in the box. The model had empty canvas and a prompt describing a product on a workbench, so it put a second product on a second workbench, at a slightly different scale, with a light source that did not agree with the first. He deleted it, went back to 256-pixel steps, and rewrote the prompt for each step to describe only what continues there (plain workshop wall continuing left, soft falloff) rather than what the image is of.

After the third step on the left side he checked the long straight edge of the bench, as the lesson says to do at every step rather than at the end, and found it had begun to curve. He added a canny control on the existing edges for the fourth step and it held. On the right side he saw a faint vertical line at the second join, visible as a slight change in contrast, and cleared it with a low-denoise pass at 0.2 over a band spanning the seam.

The last thing he did, at twenty to six, was the thing that is easy to skip: he looked at the extended image as a new picture rather than checking whether the join showed. The original frame worked because the product sat close to the right edge with tension against it. With 400 more pixels of room on that side, the tension was gone and the product floated in the middle of a wide empty bench.

The campaign was agreed at twelve images: four for print, eight for social, all sharing one visual system and one recurring character: a weaver at a pit loom, seen across a working day. The co-operative had three weeks and a small budget, and Meera was doing it alone.

She built the system file first, which is the part that pays: hard side light from camera left, 35mm at waist level, a palette of indigo, raw cotton, brass and deep brown with no green anywhere, a crop convention with the subject in the left third, and a finish of colour negative grain with lifted blacks. Then the character sheet: a contact sheet of thirty portraits, one chosen, refined with seed variation at 0.15 until the face was specific rather than average, and seven angles built from it using a face adapter at 0.8.

The decision came at the point where it always comes. Twelve images, seen side by side, four of them printed at A3. An adapter at 0.75 gets roughly seventy per cent of an identity in thirty seconds and drifts between frames. A trained LoRA gets most of the rest and costs about an hour on free Kaggle GPU hours, plus assembling and captioning the training set, plus testing epochs: call it two hours the first time.

She ran the fifteen-minute test rather than guessing: four frames with the adapter, in the four most different situations the brief contained, laid out at delivery size, shown to her flatmate who had not been in the project. The answer was immediate. The flatmate said it looked like two different women. That settled it.

Then the co-operative’s secretary made an offer that seemed helpful. One of their actual weavers, Kalaivani, has been photographed hundreds of times over the years for exhibitions and government schemes. He had a folder of about ninety images. Train it on her, he said, and the character will look like a real person because she is one.

The technical case was strong. Ninety real photographs across many years, many lightings, many angles: the opposite of the dataset fault where every image came from one afternoon against one wall. The LoRA would almost certainly be better than one trained on a character sheet.

The objection was not technical. A LoRA of a person is a tool for producing unlimited images of that person doing things she has not done, and the folder was not consent. Kalaivani had agreed, years ago, to be photographed for an exhibition. She had not agreed to have her likeness trained into a file that a designer keeps and reuses, appearing in a festival campaign she has not seen. The photographers who took the images have rights in them too, and nobody in the room knew who several of them were.

The cost of refusing was concrete: a harder, slower route to a weaker LoRA, on a three-week deadline. The cost of accepting was a permanent asset built on an agreement nobody had, in a campaign that would be printed and posted publicly, for a co-operative whose whole standing rests on its relationship with its weavers.

The artwork was finished and approved: a generated illustration of a vaccination queue at a rural health sub-centre, brand colours from the state health department, a headline in Hindi set in Inkscape over a quiet upper third. The press date was in five days. The department’s procurement note said, as such notes almost always do, artwork at 300 dpi.

A1 is 594 by 841 millimetres. At 300 dpi that is 7016 by 9933 pixels: about seventy megapixels, and a TIFF of roughly two hundred megabytes. The base render was 1216 by 832, upscaled once conservatively to about 3500 pixels on the long edge during finishing.

Getting from there to seventy megapixels is not a resize. It is a generative upscale: tiled diffusion at low denoise, forty minutes of processing, and then a careful check of every tile boundary. Ashish, who had built the poster, knew what that would do. The illustration contains a row of small figures in the queue, a signboard on the sub-centre wall, and the department’s logo composited in at the bottom. A generative upscale invents plausible detail at the scale it is asked for. Small faces become different faces. The signboard’s lettering becomes different lettering that still looks like lettering. The logo, if it went through, would come back subtly not the logo.

So his choices were three, and the deadline made them unequal.

Do the 300 dpi upscale. Cost: most of a day, a mandatory mask around the logo and the signboard so they do not go through the tiled pass, a repair pass on the queue faces afterwards, and a file large enough that the department’s email would reject it.

Deliver at 3500 pixels and say nothing. Cost: a procurement note says 300 dpi, and supplying something that does not meet a written spec without flagging it is how a job gets rejected at the printer by a person who is only checking numbers.

Ask the printer. Cost: a phone call, and the risk of an answer he did not want.

The arithmetic said the third was right before the call was made. A noticeboard poster is read from about a metre, sometimes two. The rule is roughly 3438 divided by the viewing distance in inches: at one metre, 87 dpi; at two, 44. Even with a generous safety margin at 150 dpi, A1 needs 3508 by 4967 pixels: seventeen megapixels, which is a conventional enlargement of the file he already had, with nothing invented.

There was a second problem waiting, and it was the one that would actually have caused a reprint. The state health department’s palette includes a vivid cyan-leaning blue, and Ashish had graded the whole illustration warm around it on a bright laptop. He soft-proofed in GIMP against the printer’s CMYK profile with the gamut warning on, and a large area of the sky flagged flat grey: unprintable. Converted at export without looking, that sky would have arrived on four hundred and twenty posters as a duller, greyer, slightly purple blue, and the department would have been right to say it was not their colour.