How Midjourney Credits Actually Work
Midjourney charges for GPU compute time, not per image. Every action that requires GPU processing consumes fast time from your subscription. The per-action cost is not fixed; it varies based on what you are generating.
A standard /imagine generation at default settings (4-image grid) costs approximately 1 GPU minute. Upscaling an image to a higher resolution costs additional time. Variation requests cost similar amounts to new generations. Using experimental or high-quality modes costs more than standard generation.
Fast time is the prioritized compute allocation included in your subscription plan. When fast time runs out, you either switch to relax mode (available on Standard and Pro plans, but much slower) or purchase additional fast time at roughly $4 per GPU hour. Understanding where your fast time goes is the first step to reducing how quickly it disappears.
The Highest Credit-Waste Patterns
Upscaling every image in a grid. The default workflow for many users is: generate a 4-image grid, find two or three images they like, and upscale all of them. Each upscale costs approximately the same as generating a new image. If you routinely upscale multiple images from a grid, you are spending significantly more than the base generation cost. Upscale only the single best image from each grid. If none of the four images is clearly worth upscaling, run the prompt again rather than upscaling a mediocre result.
Using redo (↺) repeatedly on the same prompt. The redo button regenerates the full 4-image grid from the same prompt with a new random seed. Each redo costs the same as the original generation. Running the same underspecified prompt multiple times hoping the output will eventually match the vision is expensive and inefficient. Write a better prompt instead.
Using high-quality upscaling modes on test generations. Midjourney has an Upscale (2x) and Upscale (4x) option, as well as the standard upscale. The high-resolution upscales cost significantly more than the standard upscale. Use the standard upscale (or no upscale) during testing and exploration. Only use high-resolution upscaling on images you have already confirmed are the ones you want to use.
Generating at --quality 2 as a default. The --quality (--q) parameter increases the render detail and coherence of the output at the cost of proportionally more GPU time. --q 2 costs approximately double the fast time of --q 1 (the default). --q 0.5 costs half. For exploration and testing, --q 0.5 or --q 1 is sufficient. Reserve --q 2 for final production runs of prompts you have already refined at lower quality.
Settings That Affect Credit Consumption
| Setting | Effect on Credit Use | Recommendation |
|---|---|---|
| --quality 0.25 | ~0.25x standard cost | Rough exploration only; low detail |
| --quality 0.5 (--q 0.5) | ~0.5x standard cost | Good for testing compositions and prompts |
| --quality 1 (default) | 1x standard cost | Standard production quality |
| --quality 2 | ~2x standard cost | Final production runs only |
| Upscale (standard) | ~1x additional cost | Use only on confirmed best image from grid |
| Upscale 2x / 4x | 2x to 4x additional cost | Use only on final selected image |
| Variation (subtle/strong) | ~1x cost per variation | Use after upscale, not before |
| Relax mode | No fast time cost | Use for low-priority or exploratory work |
A More Efficient Generation Workflow
The inefficient workflow that wastes fast time:
Write a vague prompt > Generate at --q 1 > Redo 3 times because nothing is right > Upscale two mediocre results > Run variations on both > Upscale one variation.
The above sequence spent approximately 8 to 10 GPU minutes to arrive at one usable result.
The efficient workflow:
Step 1: Write and test the prompt at --q 0.5. A --q 0.5 generation costs half the fast time and produces enough detail to evaluate whether the prompt direction is correct. Use this as your testing mode. The outputs will not look finished, but they will tell you whether the composition, subject, and mood are what you want.
Step 2: Iterate the prompt at --q 0.5 until the direction is right. Make prompt adjustments and test at low quality until you have a prompt that consistently produces the right interpretation. This phase may take 3 to 5 generations but each costs half the standard amount.
Step 3: Generate one final production run at --q 1 or --q 2. Once you have a confirmed prompt direction, run the production generation at full quality. You are now spending full fast time on a prompt you have already validated, not on an uncertain direction.
Step 4: Upscale the single best result. Select the one image from the grid that is best. Upscale it once. Done.
This workflow uses more generation steps but fewer full-quality steps. The total fast time cost is typically 30 to 50 percent lower for the same final output quality.
