Forge Academy · Practical workshop
Seeds, retries and refunds: iterate deliberately
By the InfluencerForge Editorial Team
Open the worked example ↓
TL;DR — Where the generation lane supports a seed, record it with the model, prompt and settings to make comparisons more controlled. Technical failures and filtered jobs use automatic refund handling; a completed image that misses your brief is not automatically refundable.
Repeating a prompt without diagnosing the result can spend credits without answering a useful question. In a lane that supports seeds, recording the seed alongside the model and settings helps you compare deliberate changes. Reproducibility still depends on the engine and its version.
This guide covers how reproducibility works, how to iterate deliberately, exactly when credits come back automatically, and how to budget for experimentation.
Start with the symptom
Choose the next attempt for a reason.
The face looks different
First compare the model, source references, engine, style and prompt. Remove conflicting identity descriptions. A seed cannot repair an inconsistent reference set.
The face is right, but the scene is wrong
Keep the model and settings; change one scene instruction. Reuse the recorded seed where that lane supports it, then compare the outputs side by side.
The job failed or was filtered
Check its final status and credit history before starting another job. Technical failure handling is different from disliking a successfully delivered image.
The video bends or flickers
Return to the source still if the pose is unstable. Otherwise reduce the motion request before another render.
A controlled comparison
Hold the recipe. Change one instruction.
Keep a record of the model, generation lane, resolution, aspect ratio, style, enhancement setting, prompt and seed. A seed only has meaning with those inputs; it is not a universal image identifier and should not be treated as a guarantee across engine versions.
Attempt A · baseline
Seated at a café window, light cardigan, soft daylight, medium portrait.
Attempt B · one change
Seated at a café window, light cardigan, soft daylight, wider portrait with the table visible.
Reuse the baseline's seed if available. The deliberate change is framing. Check the face and hands as well as the crop: changing text can affect more than the detail you named. These are test instructions, not fabricated before-and-after results.
What counts as a failure
Delivered and disappointing is still delivered.
For a controlled photo test, cap the session at six HD attempts: 6 × 14 = 84 credits. Stop when the experiment is no longer answering a specific question. If identity remains unstable, follow the model-validation workflow before retraining.
Run the workflow
- 1
Understand what a seed does
The seed is the starting point for a render's randomness. A seed helps reproduce a recipe when the engine, model, prompt and all settings remain compatible. It is not a guarantee across engines or model updates. Change one instruction and compare the whole result, since more than that detail may move.
- 2
Lock the seed, change one variable
When a render is almost right, reuse its seed and adjust only the element that is wrong — swap the outfit, fix the setting, tweak one phrase. Keeping the other settings fixed makes the comparison more useful, although a prompt change can still affect several details in the image.
- 3
Re-roll only when exploring
Try fresh seeds when you want to explore different compositions. If the same detail keeps failing, simplify the pose or scene and test that change instead of repeatedly buying the same attempt. Explore at 720p for 12 credits where that tier is supported, and reserve higher tiers for promising directions.
- 4
Know the refund rules
Failed or filtered jobs go through automatic refund handling; a failed model training refunds its 720-credit charge. Check the final status and credit history before retrying. A completed image you choose not to use does not qualify merely because you dislike it. If a failed charge remains, contact the Help Center with the job ID.
- 5
Budget an experiment allowance
Reserve roughly 10 to 15 percent of your monthly credits for exploration — about 300 to 450 on Starter's 3,000, or 900 to 1,350 on Creator's 9,000. If a test burns through the allowance, top-up packs never expire and stack on top of plan credits, so burst experiments do not have to distort next month's budget.
Frequently asked questions
Follow along with 100 free credits
No card required. Explore showcase models with free credits. Product-reference ads and commercial delivery require a paid plan; model training has its own eligibility and credit cost.