Banodoco Discord Database Exploration

Exploring the potential for an Open Source AI Knowledge Base

Generated: January 29, 2026

Database Overview

1,046,692
Discord Messages
6,624
Community Members
227
Channels
463
Daily Summaries

Date Range Coverage

Discord Messages

August 8, 2023 to January 29, 2026

~29 months of continuous data coverage

Message Volume by Period

Period Messages Distribution
Aug - Dec 2023 124,595
Jan - Dec 2024 262,808
Jan - Jun 2025 273,282
Jul - Dec 2025 319,886
Jan 2026 66,053
Complete Coverage: The database now has continuous message history from August 2023 to present, including the full 2024 period covering FLUX, Stable Diffusion 3, CogVideoX, and early HunyuanVideo discussions. Gap was resolved by @pom in January 2026.

AI-Generated Daily Summaries

November 4, 2025 to January 29, 2026

~87 days of summaries (463 total)

Summary Coverage: Daily summaries only cover Nov 2025 - Jan 2026 (~87 days). The database has ~660K messages from Aug 2023 - Oct 2025 that exist but don't have AI-generated summaries yet.

Channels with Daily Summaries

21 channels currently have AI-generated daily summaries:

wan_chatter wan_comfyui wan_training wan_gens wan_resources ltx_chatter ltx_training ltx_resources ltx_gens hunyuanvideo flux comfyui chroma trellis kandinsky-5 qwen-image sora z-image newbie daily_summaries main_test

Sample Daily Summaries

These AI-generated summaries show the quality and depth of technical knowledge being captured:

LTX Chatter - January 24, 2026

📨 1088 messages sent

  • Kijai released major memory optimizations for LTX-2 including in-place NAG application and FFN chunking, enabling ~2000 frame renders with only 18GB VRAM
  • Community discovered dev model requires distill LoRA at 0.2-0.6 strength to avoid noisy outputs at high resolutions
  • sa_solver sampler runs nearly 2x faster than res_2s (79s vs 149s) with negligible quality difference

Flux - January 24, 2026

📨 100 messages sent

  • Kijai discovered Flux2's tiny VAE (TAEF2) delivers surprisingly high quality despite small size
  • Dever successfully trained Arcane style LoRA using unpaired text-to-image training
  • koshi hacked Deforum to work with native Klein 4B for hybrid video generation

LTX Training - January 24, 2026

📨 204 messages sent

  • crinklypaper fixed high-pitched audio in LTX-2 LoRA by matching dataset FPS to training settings - resolved in just 500 additional steps
  • MOV revealed I2V/T2V training requires adding dataset twice - training improved drastically after implementing this
  • Alisson Pereira trained IC LoRA for face swapping using 150 video pairs

Database Schema Highlights

Core Tables

Table Description Key Fields
discord_messages All chat messages content, author_id, channel_id, attachments, embeds, created_at
discord_channels Channel metadata channel_name, category_id, description
discord_members User profiles username, global_name, avatar_url, twitter_handle
daily_summaries AI-generated summaries full_summary (JSON), short_summary, date, channel_id

Summary Structure

Each daily_summaries.full_summary contains structured JSON with:

  • title - Topic headline
  • mainText - Overview with user attribution
  • message_id - Link back to source message
  • subTopics - Array of related discussion points with media references

All Channel Categories

Video Generation Models

wan_chatter wan_comfyui wan_training hunyuanvideo cogvideox mochi ltx_chatter animatediff stable-video-diffusion cosmos step-video sora veo3

Image Generation

flux sdxl-turbo stable-cascade omnigen qwen-image lumina-image-2 chroma hidream

Training & Fine-tuning

training wan_training flux_training hunyuanvideo_training cogvideox_training ltx_training training_control_loras

Technical Infrastructure

comfyui nodes ml_engineering ml_learning ml_papers ai-powered-coding

3D & Motion

trellis hunyuan3d 2d-3d-traditional liveportrait champ aniportrait

Audio

mmaudio ace-step music yue zonos

Opportunities for Knowledge Base

1. Pre-Generated Summaries as Foundation

The 429 daily summaries are already structured with topic titles, key contributors, and source message links. These can serve as the backbone of a knowledge base without additional processing.

2. Rich Technical Content

Based on the sample summaries, the content includes:

  • Specific VRAM requirements and optimization techniques
  • Model configuration parameters (LoRA strengths, samplers, steps)
  • Troubleshooting solutions with attribution to community experts
  • Training recipes with dataset sizes and settings

3. Expert Attribution

Summaries consistently attribute discoveries to specific community members (Kijai, crinklypaper, MOV, etc.), making it possible to identify and highlight key contributors and their areas of expertise.

Opportunity: Historical Summaries

Summary generation opportunity:

  • ~660K messages from Aug 2023 - Oct 2025 exist in the database but don't have AI-generated summaries yet
  • This includes major eras: AnimateDiff, FLUX launch, SD3, CogVideoX, early HunyuanVideo

Options for the knowledge base:

  • Generate retrospective summaries for historical content
  • Focus initial knowledge base on Nov 2025+ (already summarized)
  • Prioritize high-activity channels for retrospective processing

Next Steps

  1. Deep dive into summary structure - Explore the full JSON schema of daily_summaries to understand all available fields
  2. Analyze channel coverage - Determine which channels have the most valuable technical content
  3. Build export pipeline - Create scripts to extract and transform data for the knowledge base
  4. Design navigation - Plan how users will browse by model, topic, contributor, or date
  5. Generate retrospective summaries - Process Aug 2023 - Oct 2025 messages for historical coverage