FLIR thermal imaging of sprinter mid-stride, heat gradient mapping from deep cobalt (extremities, 25°C) through emerald green (torso, 33°C) to burnt orange core (38°C), extreme radial motion blur on limbs suggesting 10m/s velocity, heavy 35mm film grain with push-processed density, analog CRT phosphor persistence with faint horizontal scan lines, absolute matte black background (0.001 lux), single 1px neon cyan horizon line at lower third, infrared thermography with 0.05°C sensitivity, atmospheric sensor noise, gaussian bloom on >35°C zones, vintage 1980s scientific documentation aesthetic, thermal crossover artifacts at joints --ar 1:1 --style raw --s 250
Quick Tip: Click the prompt box above to select it, then press Ctrl+C (Cmd+C on Mac) to copy. Paste directly into Midjourney, DALL-E, or Stable Diffusion and customize to make it your own!
Two thermal renders of the same sprinter can look nothing alike. The naive version treats heat as decoration: a body silhouette filled with a rainbow gradient, uniform saturation from head to heel, no anatomical logic. The informed version maps temperature as data. Extremities sit cool at roughly 25°C and read deep cobalt; the torso warms to about 33°C and shifts to emerald; the core concentrates at 38°C in burnt orange. The naive image is a colored person. The informed image is a measurement.
The gap between those two outcomes is not artistic taste. It is prompt architecture. Thermal imaging is a rendering of quantitative sensor data, and diffusion models respond to numbers in ways they do not respond to adjectives. Writing "hot colors" produces a palette. Writing "38°C core, 25°C extremities" produces a gradient topology. This article compares three concrete approaches to thermal prompting, each with a distinct visible signature, so you can choose the one that matches your production goal.
Approach One: Flat False-Color Mapping
The simplest approach specifies only a palette: "thermal camera image, rainbow gradient, blue to red body." The mechanism is straightforward. The model samples color distribution from its training data of thermal photography and applies it uniformly across the subject. Every pixel of the silhouette receives some saturation value because the prompt never told it where heat should concentrate.
The visible signature is unmistakable. Colors are vivid and evenly distributed, with no true cold zones. Backgrounds often pick up faint color bleed because the model has no instruction to hold them at zero. Edges are soft and the figure reads as a sticker pasted onto the scene. This approach is fast and works for thumbnails or abstract graphic work where anatomical accuracy is irrelevant.
The trade-off is credibility. Without a temperature range, the model cannot distinguish a hand from a chest. Any viewer who has seen real thermography will register the flatness immediately, even without naming why. This is the same failure mode that flattens graphic art prompts when color is specified but spatial logic is not.
Approach Two: Temperature-Anchored Gradient Mapping
The second approach replaces palette language with a numeric scale. The prompt assigns specific °C values to body zones and names the corresponding colors: deep cobalt at 25°C for extremities, emerald green at 33°C for torso, burnt orange at 38°C for core. This works because the model has absorbed the visual conventions of FLIR and infrared thermography, where temperature and hue are locked in a known relationship.
Once the scale is defined, the model distributes heat anatomically. Shoulders and chest warm first, hands and feet cool last, and joints show crossover zones where warm and cool regions meet. The signature is a gradient that follows physiology rather than geometry. Saturation is no longer uniform; it is a map. Backgrounds stay near black because the prompt's coldest value anchors to the subject's coldest surfaces.
The trade-off is control cost. Every zone you specify adds a constraint the model must satisfy simultaneously. Push too many temperature brackets into one prompt and the gradient collapses into noise or reverts to flat color. The practical ceiling sits around four to five distinct temperature bands before coherence degrades. This constraint is worth understanding before layering additional effects.
Approach Three: Sensor Realism with Analog Artifacts
The third approach treats the thermal render as a recording from a specific instrument, not a clean data visualization. It layers sensor noise, 35mm film grain with push-processed density, CRT phosphor persistence, and faint horizontal scan lines over the temperature map. The mechanism is perceptual: real thermography carries the fingerprints of its capture chain, and the model reproduces those fingerprints when they are named.
The visible signature is a loss of digital cleanliness. Grain breaks up smooth gradients, bloom appears around zones above roughly 35°C, and scan lines give the image a 1980s scientific documentation texture. Thermal crossover artifacts at joints become visible as small discontinuities where the gradient cannot resolve smoothly. The result reads as evidence rather than illustration.
This approach compounds the cost of the previous one. Sensor realism depends on the temperature map already being correct; grain over a flat false-color fill produces noise, not credibility. The correct build order is scale first, artifacts second. For renders that need to survive close inspection, the same principle applies as in feathered portrait work: the underlying structure carries the image, and surface treatment only amplifies what is already there.
Motion Blur as a Velocity Signal
Static thermal images of a runner look like a standing figure with warm legs. To read as mid-stride at speed, the prompt needs a velocity parameter expressed through rendering behavior. Radial motion blur on the limbs, described as consistent with roughly 10 m/s, tells the model to smear the extremities along their direction of travel while keeping the torso relatively resolved.
The physics logic matters here. Fast-moving surfaces lose thermal detail in real infrared capture because the sensor integrates over a short exposure window. The model reproduces this if the prompt connects blur to speed rather than treating it as a stylistic filter. Blur applied uniformly across the whole figure reads as camera shake; blur concentrated on limbs and absent from the core reads as motion.
Pairing the blur with a hard black background at near-zero lux removes competing luminance and lets the gradient do the work. A single thin cyan horizon line at the lower third provides ground reference without introducing terrain detail that would pull attention from the heat map. Every element either supports the temperature read or is removed.
The Boundary Signal
Temperature-anchored thermal prompting stops working when the subject's heat distribution stops being legible. The first signal is gradient inversion: warm zones appearing at extremities while the core stays cool. That indicates the model has lost the anatomical anchor and is distributing color by spatial randomness instead of physiology. The second signal is band collapse, where four or more temperature brackets merge into a single mid-tone.
Both failures share a cause. The prompt has exceeded the number of simultaneous constraints the model can hold in a coherent spatial map. The remedy is subtraction, not addition: remove the least important effect layer, typically the sensor artifacts, and re-render with the temperature scale intact. Once the gradient resolves cleanly again, artifacts can be reintroduced one at a time until the boundary reappears.
Working with tools such as Midjourney or Leonardo, that boundary sits in a predictable place for most thermal subjects: three to five temperature zones, one motion treatment, and one background constraint. Beyond that count, the model stops measuring heat and starts approximating it. The moment you can no longer point to which zone is 25°C and which is 38°C in the output, the prompt has crossed the line.
Label: Cinematic
Key Principle: Anchor thermal prompts to a temperature scale, not color adjectives. Naming exact °C per body zone forces the model to distribute its heat map anatomically instead of flooding the whole silhouette.
Comments
Post a Comment