At a glance
| Item | Value |
|---|---|
| Video | Civitai MiniMax H3 (Comfy), standard 30 steps — 35 blue per second (4 s 140 · 10 s 350) |
| Output resolution | 9:16 704×1280 (720p class). Resolution does not change the price |
| Image | ChatGPT image on the free plan — one reference is enough |
| Blue buzz you can earn yourself per day | Fixed 220 (five items) + daily challenge entry prize (set per challenge, recently 200) |
| Daily reset | 00:00 UTC |
| What a free account can generate | PG / PG-13 |
Video prices exclude the creator tip. If the generator adds a tip to the total, remove it to compare. Civitai can change rewards and prices at any time.
About Civitai.com
Civitai is a site for sharing image and video generation models (checkpoints and LoRAs), and it also runs a web generator. No install or GPU — you generate in the browser: civitai.com
| Item | What it is |
|---|---|
| Buzz | The on-site currency for generation |
| Blue buzz | Earned free through daily activity → spent on generation. What this guide covers |
| Yellow buzz | Bought with money. Not used in this guide |
| What a free account can generate | PG / PG-13 (beyond that needs a membership) |
| Your reward progress | My Buzz Dashboard — civitai.com/user/buzz-dashboard (progress per reward and time to reset) |
Recommended workflow · GPT image → one 9-second Civitai H3 clip
- Make one character reference on ChatGPT (free plan) — vertical 9:16, and describe the person, outfit, prop, room and light
- Civitai generator → Video → MiniMax H3 → version H3 (Comfy), upload the reference
- Aspect 9:16, Duration 9 s, Turbo off (standard 30 steps)
- In the prompt, split the scenes as <Shot 1> 00:00–00:03 · <Shot 2> 00:03–00:06 · <Shot 3> 00:06–00:09 and give each shot its own framing, action and line
- Don’t leave the outfit to the reference alone — write it out (e.g. cream cable-knit cardigan … Do not change the cardigan color)
- Generate once (315 blue) → in an editor (CapCut or similar), only fine-tune the cut points and dead air
| Shot | Role | Line (example, Korean) | Camera and action |
|---|---|---|---|
| <Shot 1> 0–3 s | Hook | "This video cost me nothing." | Medium close-up, leans toward the lens and talks right away |
| <Shot 2> 3–6 s | Reset | "I made it with free daily points." | Hard cut → tighter shot, glances at the monitor and turns back |
| <Shot 3> 6–9 s | CTA | "How-to is in my profile link." | Hard cut → seated full shot, points down once |
# Trim cut points and dead air from the single 9-second clip (seconds; adjust while watching)
ffmpeg -i nari_9s.mp4 -filter_complex "\
[0:v]trim=0:3.25,setpts=PTS-STARTPTS[v0];[0:a]atrim=0:3.25,asetpts=PTS-STARTPTS[a0];\
[0:v]trim=3.95:6.85,setpts=PTS-STARTPTS[v1];[0:a]atrim=3.95:6.85,asetpts=PTS-STARTPTS[a1];\
[0:v]trim=6.9:9.4,setpts=PTS-STARTPTS[v2];[0:a]atrim=6.9:9.4,asetpts=PTS-STARTPTS[a2];\
[v0][a0][v1][a1][v2][a2]concat=n=3:v=1:a=1[v][a]" \
-map "[v]" -map "[a]" -c:v libx264 -crf 18 -c:a aac nari_cut.mp4The full prompt is "NARI 9-second 3-shot" in the prompt kit. Making three separate 4-second clips cost 420 blue and the speech started anywhere from 0.1 to 2.8 s into each clip. One 9-second clip is cheaper and easier to edit.
Tested · one 9-second generation
One NARI reference (941×1672) made with ChatGPT image, H3 (Comfy) standard 30 steps · 9 s · 704×1280, generated twice.
| Item | Run 1 (no outfit sentence) | Run 2 (outfit sentence added) |
|---|---|---|
| Cost · gen time | 315 blue · 449.1 s | 315 blue · 433.5 s |
| Output length | 9.42 s | 9.42 s |
| Hard cuts at | 3.50 s · 7.13 s | 3.29 s · 6.88 s |
| Speech heard at | 0.2–3.3 / 4.5–6.6 / 7.2–9.2 s | 0.2–3.0 / 4.4–6.4 / 7.3–9.3 s |
| 3 lines, speech recognition | all matched | all matched |
| Outfit | cream cardigan turned navy | kept as in the reference |
- Both runs followed the prompt: three shots, two hard cuts, and each line inside its own shot.
- Cuts land 0.1–1.1 s off the requested 3 s and 6 s. Trim that and any leading silence (about 1.1 s before the run-2 reset line) in the editor. Run 2 trimmed with the command above came to 8.67 s with the dialogue still matching.
- With only the reference, the outfit color can change. Writing the outfit out and adding "do not change" kept it.
Daily blue buzz · fixed 220 + daily challenge
| Reward | How | Each | Daily cap |
|---|---|---|---|
| Daily Boost | Claim Daily Boost in the generator | 25 | 25 |
| First post | Post one image | 25 | 25 |
| Reactions | React to other people’s content | 2 | 100 (50×) |
| Follows | The first 3 people you follow that day | 10 | 30 |
| Generator feedback | Rate your generator results (like/dislike) | 4 | 40 (10×) |
| Total | 220 |
The daily challenge entry prize is not a reward with a daily cap. It is a prize Civitai sets for each challenge, so how much you get and how much you keep depends on that day’s challenge.
| Daily challenge | Details |
|---|---|
| Prize · entries needed | Set per challenge (Entry Prize on the challenge page). The last 41 system challenges (2026-08-28 to 10-07) were all 10 accepted entries → 200 |
| What counts | Only images made in the Civitai generator with that day’s required model (LoRA). Images made elsewhere with the resource credited by hand are rejected |
| Payout | Paid automatically before the deadline once enough entries are accepted (ACCEPTED). Accepted entries count, not submissions |
| Generation cost | About 30 blue for 10 images on SDXL/Pony days → about 170 left. If the only required model is Flux or Qwen, 10 images cost 230–330 blue, a loss (none of the last 41) |
| Deadline | Switches to the next challenge every day at 00:00 UTC |
- Up to 420 on a day you complete the challenge (about 390 after generation cost); 220 on a day you skip it or fall short of accepted entries.
- The 40 generator-feedback buzz needs ratings on your own generations. Rating your 10 challenge images covers it.
- Entries are moderated. Photoreal people who look underage get removed as a violation, so describe people as adults.
- User-run challenges with an entry fee may have no entry prize. Pick system challenges with a 0 fee.
- There is no ad-watching reward. It is defined in the source but switched off, and it does not appear in the My Buzz Dashboard reward list (checked 2026-10-02).
You also earn when other people engage with your content (daily caps apply).
| What others do | Each | Daily cap |
|---|---|---|
| React to your content (last 30 days) | 2 | 100 |
| Add your content to a collection | 2 | 100 |
| Remix accepted into your gallery | 20 | 100 |
| Sticker accepted on your content | 10 | 100 |
One-time: a 100 sign-up bonus (we confirmed it does not repeat the next day), and 500 each for the new user and the referrer when someone joins through a referral link.
What not to do
- Farming buzz with multiple accounts — violates Terms 11.8.
- Repeating reactions or follows with bots, scripts or automation — violates Terms 11.9.
- It is enforced: a September 2026 notice reported 385 buzz-farming/multi-accounts and an 8,563-account reaction-bot chain banned.
- False reports for the reward — the 50-buzz report reward pays only on accepted reports, and abuse gets accounts banned.
Everything in this guide assumes one account, used by a person.
H3 cost by setting
| Setting | Blue |
|---|---|
| H3 (Comfy) standard 30 steps · 4 s | 140 |
| H3 (Comfy) standard · 6 / 10 / 15 s | 210 / 350 / 525 |
| H3 (Comfy) Turbo 8 steps · 4 / 6 s | 48 / 72 |
| H3 (API) 2K · 6 s | 1,020 |
Turbo is cheap but weak on motion. We compared three Turbo 4-second clips.
| Turbo input | Result |
|---|---|
| Portrait reference + dialogue (reference mode) | 52.6 s · all 10 words matched · no cuts · framed a little tighter than the source |
| Same image as the first frame | 42.1 s · dialogue matched · cut to a closer shot within 0.5 s |
| Scene with big hand gestures | 63.1 s · all 11 words matched · ghosting on fast hand motion |
Use standard 30 steps for what you publish and Turbo for trying prompts. Reference mode held the framing better than first-frame mode.
Images · ChatGPT free, or Qwen 2.1 inside Civitai
ChatGPT’s free plan can generate images with a usage limit; OpenAI does not publish the number (when you hit it, the message tells you when you can try again). This workflow needs only one reference image, so the free plan is enough.
To make images inside Civitai too, use Qwen 2.1 (spends blue).
| Qwen 2.1 setting | Blue |
|---|---|
| 1K 9:16 (576×1024) · 25 steps | 5 |
| 1K 1:1 (1024×1024) · 25 steps | 8 |
| 2K 9:16 (1152×2048) | 18 |
Our seven Qwen 2.1 images finished in 18–57 s with exact Korean text. English text only avoided a white band and clipped letters when we said "full-bleed … no white border or band".
Picks · tested value
| Use | Pick | Cost | Why |
|---|---|---|---|
| Video you publish | H3 (Comfy) standard 30 steps | 35 blue per second | Korean dialogue and audio, consistent person from one reference |
| Trying prompts | H3 (Comfy) Turbo | 12 blue per second | About a minute per clip, ghosting on big motion |
| Reference image | ChatGPT image | free (limited) | Keeps all your blue for video |
| Images inside Civitai | Qwen 2.1 · 1K 9:16 | 5 blue | Accurate Korean text |
| Skip | H3 (API) | 1,020 blue for 6 s | About 5× standard 6 s (210) |
Prompt kit
The 9-second three-shot prompt and formula (hook, reset, CTA), the ChatGPT reference request and the image/video formulas are in the resource library — search "H3", "NARI" or "hook" there: resary.com/en/resources/civitai-free-prompt-kit
Sources
| Source | Note |
|---|---|
| Civitai Terms of Service — civitai.com/content/tos | 11.4, 11.8, 11.9, 14.2 |
| Civitai reward definitions (public source) — github.com/civitai/civitai/tree/main/src/server/rewards | amounts and caps as of commit 2c18205 (2026-10-01) |
| Bot cleanup results — civitai.com/articles/35068/re-august-bot-cleanup-results | 2026-09-09, banned account counts |
| Buzz changes — civitai.com/articles/21872/buzz-changes | free accounts limited to PG/PG-13 |