Home Creators Posts Import Register Favorites Logout

Videos

  • 764986651.mp4

Downloads

Content

Hey!

You've seen the video: Fish Audio Drama 3 (https://www.patreon.com/aifuturetech/posts/fish-audio-drama-172471837) cloned two reusable character voices, performed a couple's argument with plain-English emotion tags, and MiniMax H3 animated it with the matching character sheets. This post is the run guide — what to download, what to wire, what to type, and the exact settings that survived my testing.

The free TTS workflow from the video stays free. This post carries the two H3 sampling setups (base run + the upscale pass), the five-part animated script, and the voice/audio prep files.

----------------------------------------

WHAT TO DOWNLOAD (MODEL FILES)

----------------------------------------

Drama 3 TTS side — from https://huggingface.co/PedroMarinhoDev/Drama-3-ComfyUI

- drama3_int8experts_bf16.safetensors (~38.7 GB) → models/diffusion_models (the one I ran — same accuracy as FP8, 14–17% faster, no silent takes in my 12-clip test)

- drama3_fp8experts_bf16.safetensors (~38.6 GB) → models/diffusion_models (author default; swap in if you prefer the original bytes)

- drama3_codec.safetensors (~1.6 GB) → models/vae (always required)

- qwen3_asr_1.7b_bf16.safetensors (~4.1 GB) → models/asr (create this folder — ComfyUI doesn't ship it)

H3 side — Comfy-Org MiniMax H3 weights + your usual LoRAs (all under models/, minimax-h3 subfolder in mine)

- Minimax-h3_Singularity_ref2va_v1.3_int8.safetensors → models/diffusion_models

- qwen3vl_32b_minimax_h3_int8_convrot.comfy.safetensors → models/text_encoders (loaded via CLIPLoader, type "minimax")

- minimax_h3_video_vae_fp16.safetensors / minimax_h3_audio_vae_fp32.safetensors → models/vae

- minimax_h3_ref2v_turbo_4step_v0.1_comfyui_bf16.safetensors → models/loras

- h3-realism-people-t2v-i2v-r2v(r34l1sm).safetensors → models/loras

- minimax_h3_lms_v1.0_r64.safetensors → models/loras (2-sampling workflow only)

- minimax_h3_t2va_veda_8nfe_600step_preview_fp8.safetensors → models/veda

- minimax_h3_latent_upscaler_3d_conv_v1_fp16.safetensors → models/upscale_models (2-sampling workflow)

Custom nodes: ComfyUI-Drama3 (https://github.com/PedroMarinhoDev/ComfyUI-Drama3 — clone into custom_nodes, no pip installs) plus your existing H3 / VideoHelperSuite / KJNodes / DLSS packs.

----------------------------------------

FILES ATTACHED TO THIS POST

----------------------------------------

- drama3_text_to_speech.json (also free — the TTS + voice-save graph)

- MinimaxH3_drama3-1sampling-20261015.json

- MinimaxH3_drama3-2Sampling-20261015.json (base sampling → latent upscale → sa_solver refine, the sharper pass)

- drama3_h3_prompts.md (five 15s Ref2VA prompt blocks, one per argument beat)

- seg1–seg5 .wav clips (the 58s argument cut at natural pauses — drop-in reference audio)

----------------------------------------

QUICK START

----------------------------------------

1. Drag in drama3_text_to_speech.json. Point the two VHS_LoadAudio paths at your voice clips (audio file).

2. Queue once with the Save Drama Voice nodes unmuted — that bakes each voice to output/drama3_voices/ as a tiny safetensors. After that, mute Load Audio + Qwen3-ASR + Encode Voice entirely and run from Drama3LoadVoice — it saves real memory per run.

3. Write your dialogue in Drama 3 Generate, queue, listen. Fix the voice script, not the settings, when a take misbehaves.

4. Export the finished speech. For the animation: cut it into 13–14 s segments at pauses (the five seg wavs attached are mine), so every clip lands inside H3's 15 s budget.

5. Drag in one of the MinimaxH3_drama3 files. Re-point the image/audio loaders at your character sheets and your segment wav. Paste that segment's Ref2VA block from drama3_h3_prompts.md into the prompt string node. Queue. That's the whole loop.

----------------------------------------

WHAT TO INPUT (PROMPT FORMATS)

----------------------------------------

TTS side — Drama 3 reads directions as plain words, no fixed tag list:

- Start every spoken line with its speaker: <|speaker:0|> ... <|speaker:1|> — numbers count the voice sockets on Drama 3 Generate, not names.

- Emotion per line goes in a bracket prefix: [tight, controlled], [defensive], [laughs bitterly], [with heavy, deflated sadness]

- Span control: <slow>...</slow>, <whisper>...</whisper> wrap the words they shape; [pause], [sighs], [clears throat] land at their position.

- NEVER start a line with "..." — the runtime swallows a leading ellipsis and the whole line vanishes (this bit me on the first take; it skipped three lines). Use [pause] instead. Mid-line ellipsis is safe.

- Keep exchanges to a handful of lines per run — long multi-line arguments drift.

H3 side — Ref2VA six-section block, per segment (all five pre-written in the attached md):

subject_definitions / summary / retention_analysis / detailed_description / overall_soundscape / non_diegetic_music

Wiring convention the prompts assume: Picture 1 = female sheet, Picture 2 = male sheet, Picture 3 = scene plate, Ref Audio 0/1 = the two seed voices (timbre), Audio 1 = the segment wav (dialogue, fully_copy). Keep S1 = female, S2 = male in every segment so stitched clips don't swap voices.

----------------------------------------

KEY SETTINGS (AND WHY)

----------------------------------------

Drama 3 Generate:

- temperature 1.0, top_p 0.9, top_k 30 — Fish Audio's own server defaults; lower values ran flat and silent for me

- max_frames 600 (~28 s) shipped; raise per clip length, split_long_text stays on

- Drama3Loader: I ran experts "in VRAM" on my card; on anything under ~24 GB use "in RAM" with expert_cache_gb 0 — the offload path is why this runs on consumer GPUs

H3 base sampling (both workflows):

- Duration formula node: max(5, round(a*24)) + pad-to-17n+5 — feed the segment length (a) in seconds; H3 needs the 17n+5 frame grid (24 fps)

- Resolution 1344x768, reference length 124, "match" on MiniMaxH3ReferenceToVideo — the two-person framing that kept both sheets in range

- Sigma shift 12/3, chunk feed-forward 2/4096, torch settings on: leave as shipped

- LoRAs all at strength 1.0: ref2v_turbo_4step + realism-people(r34l1sm) + character_swap — the trigger string "r34l1sm" is prefixed by the concatenate node, keep it or the realism LoRA sleeps

- Veda sparse attention 32/32 on, DLSS 1.5x Quality — the speed stack; drop Veda if a segment has fast motion, it morphs on action

- Seed 43 fixed while iterating prompts; randomize only when the clip is approved

2-sampling refine (paid workflow, second sampler chain):

- sa_solver + BasicScheduler simple/1/0.2 + LMS r64 LoRA — the sharpening pass, same seed 777 fixed

- MMH3LatentUpscaleWithModelParams 1280x704 cuda/bf16 + spatial split 864x480 — this is the piece that pulls faces and hands back up; one pass costs time, keep it for keepers only

- VAEDecodeTiled 512 — decode headroom on the upscaled latents

----------------------------------------

TIPS

----------------------------------------

1. Cut segment audio at the gaps, not at 15.000 — the attached segs end on clean pauses; a clip that starts mid-word makes H3's lip-sync drift from frame one.

2. Character sheets as turnaround sheets work fine, but the model locks the outfit — if your next video changes wardrobe, re-feed sheets and update the subject_definitions text with it.

3. Save voices with the subfolder prefix (drama3\Name) — the node writes into output/drama3_voices and a bare name clutters it.

4. The Ref Audio sockets carry timbre; the segment wav carries the actual words and delivery. Wire both — timbre-only reference lost Anika's accent in my test.

5. The laugh at the end of segment 5 is the hardest tag in the whole set — quantized and even full-precision packs occasionally swallow dialogue after laughter. If it drops, re-queue that segment alone before editing anything.

----------------------------------------

COMMON ISSUES

----------------------------------------

Lines disappear from the TTS take → leading "..." on the line → remove it, use [pause] before the bracketed delivery note.

Take goes silent after a second or two → voice transcript doesn't cover the clip word-for-word, or the clip is cut mid-sentence → re-transcribe with Qwen3-ASR, trim the source clip between sentences, re-queue for a new seed.

H3 output saves as vid_000XX.png instead of mp4 → save_output off on the first VHS_VideoCombine in the 2-sampling file (it's the preview pass; the final combine at node 6651 writes the video) → or just point the MP4 combiner at the refine output.

Voices swapped between segments → seed audio sockets re-ordered between runs → Ref Audio 0 is always the female, Ref Audio 1 the male, and Audio 1's S1/S2 order follows who speaks first in the clip.

Realism LoRA does nothing → prompt missing "r34l1sm" → keep the concatenate node wired; it injects the trigger in front of every prompt.

Out of memory on the refine pass → run the upscale chain alone: mute the base sampling combine, or drop spatial split tile to 640x352.

----------------------------------------

Any trouble running it, drop a comment with your GPU, VRAM, and the exact error and I'll look at it personally.

Thanks for supporting the channel. You guys make these deep dives possible.

— Benji

Files

Comments

No comments found for this post.